wx44wx/pipecat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'main' into groundingMetadata

Browse Source
This commit is contained in:
Pete
2025-06-30 19:48:55 -04:00
committed by GitHub
parent 1cf0b35ac1 224d2cedc8
commit a297e4208e
148 changed files with 9172 additions and 1291 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
CHANGELOG.md
View File

@@ -9,25 +9,48 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
### Added
- Added `watchdog_coroutine()`. This is a watchdog helper for couroutines. So,
if you have a coroutine that is waiting for a result and that takes a long
time, you will need to wrap it with `watchdog_coroutine()` so the watchdog
timers are reset regularly.
### Fixed
- Fixed a `AWSNovaSonicLLMService` issue introduced in 0.0.72.
## [0.0.73] - 2025-06-26
### Fixed
- Fixed an issue introduced in 0.0.72 that would cause `ElevenLabsTTSService`,
`GladiaSTTService`, `NeuphonicTTSService` and `OpenAIRealtimeBetaLLMService`
to throw an error.
## [0.0.72] - 2025-06-26
### Added
- Added logging and improved error handling to help diagnose and prevent potential
Pipeline freezes.
- Added `WatchdogQueue`, `WatchdogPriorityQueue`, `WatchdogEvent` and
`WatchdogAsyncIterator`. These helper utilities reset watchdog timers
appropriately before they expire. When watchdog timers are disabled, the
utilities behave as standard counterparts without side effects.
- Introduce task watchdog timers. Watchdog timers are used to detect if a
Pipecat task is taking longer than expected (by default 5 seconds). It is
Pipecat task is taking longer than expected (by default 5 seconds). Watchdog
timers are disabled by default and can be enabled globally by passing
`enable_watchdog_timers` argument to `PipelineTask` constructor. It is
possible to change the default watchdog timer timeout by using the
`watchdog_timeout` constructor argument when creating a `PipelineTask`. With
watchdog timers it is also possible to log how long each processing step is
taking (e.g. processing an element from a queue inside a task). This is done
with the `enable_watchdog_logging` constructor argument when creating a
`PipelineTask.` It is also possible to control these two values per each frame
processor. That is, you can set set `enable_watchdog_logging` and
`watchdog_timeout` argument. You can also log how long it takes to reset the
watchdog timers which is done with the `enable_watchdog_logging`. You can
control all these settings per each frame processor or even per task. That is,
you can set `enable_watchdog_timers`, `enable_watchdog_logging` and
`watchdog_timeout` when creating any frame processor through their constructor
arguments. Finally, you can also set these values per task. So, if you are
writing a frame processor that creates multiple tasks and you only want to
enable logging for one of them, you can do so by passing the same argument
names to the `FrameProcessor.create_task()` function. Note that watchdog
timers only work with Pipecat tasks but not if you use `asycio.create_task()`
or similar.
arguments or when you create a task with `FrameProcessor.create_task()`. Note
that watchdog timers only work with Pipecat tasks and will not work if you use
`asycio.create_task()` or similar.
- Added `lexicon_names` parameter to `AWSPollyTTSService.InputParams`.
@@ -84,6 +107,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
### Fixed
- Fixed an issue that would cause heartbeat frames to be sent before processors
were started.
- Fixed an event loop blocking issue when using `SentryMetrics`.
- Fixed an issue in `FastAPIWebsocketClient` to ensure proper disconnection

109
CONTRIBUTING.md
View File

@@ -41,36 +41,107 @@ We use Ruff for code linting and formatting. Please ensure your code passes all
We follow Google-style docstrings with these specific conventions:
- Class docstrings should fully document all parameters used in `__init__`
- We don't require separate docstrings for `__init__` methods when parameters are documented in the class docstring
- Property methods should have docstrings explaining their purpose and return value
**Regular Classes:**
Example of correctly documented class:
- Class docstring describes the class purpose and key functionality
- `__init__` method has its own docstring with complete `Args:` section documenting all parameters
- All public methods must have docstrings with `Args:` and `Returns:` sections as appropriate
**Dataclasses:**
- Class docstring describes the purpose and documents all fields in a `Parameters:` section
- No `__init__` docstring (auto-generated)
**Properties:**
- Must have docstrings with `Returns:` section
**Abstract Methods:**
- Must have docstrings explaining what subclasses should implement
**`__init__.py` Files:**
- **Skip docstrings** for pure import/re-export modules
- **Add brief docstrings** for top-level packages or those with initialization logic
**Enums:**
- Class docstring describes the enumeration purpose
- Use `Parameters:` section to document each enum value and its meaning
- No `__init__` docstring (Enums don't have custom constructors)
#### Examples:
```python
class MyClass:
"""Class description.
# Regular class
class MyService(BaseService):
"""Description of what the service does.
Additional details about the class.
Args:
param1: Description of first parameter.
param2: Description of second parameter.
Provides detailed explanation of the service's functionality,
key features, and usage patterns.
"""
def __init__(self, param1, param2):
# No docstring required here as parameters are documented above
self.param1 = param1
self.param2 = param2
def __init__(self, param1: str, param2: bool = True, **kwargs):
"""Initialize the service.
Args:
param1: Description of param1.
param2: Description of param2. Defaults to True.
**kwargs: Additional arguments passed to parent.
"""
super().__init__(**kwargs)
@property
def some_property(self) -> str:
"""Get the formatted property value.
def sample_rate(self) -> int:
"""Get the current sample rate.
Returns:
A string representation of the property.
The sample rate in Hz.
"""
return f"Property: {self.param1}"
return self._sample_rate
async def process_data(self, data: str) -> bool:
"""Process the provided data.
Args:
data: The data to process.
Returns:
True if processing succeeded.
"""
pass
# Dataclass
@dataclass
class ConfigParams:
"""Configuration parameters for the service.
Parameters:
host: The host address.
port: The port number. Defaults to 8080.
timeout: Connection timeout in seconds.
"""
host: str
port: int = 8080
timeout: float = 30.0
# Enum class
class Status(Enum):
"""Status codes for processing operations.
Parameters:
PENDING: Operation is queued but not started.
RUNNING: Operation is currently in progress.
COMPLETED: Operation finished successfully.
FAILED: Operation encountered an error.
"""
PENDING = "pending"
RUNNING = "running"
COMPLETED = "completed"
FAILED = "failed"
```
# Contributor Covenant Code of Conduct

36
docs/api/conf.py
View File

@@ -1,5 +1,6 @@
import logging
import sys
from datetime import datetime
from pathlib import Path
# Configure logging
@@ -13,7 +14,8 @@ sys.path.insert(0, str(project_root / "src"))
# Project information
project = "pipecat-ai"
copyright = "2024, Daily"
current_year = datetime.now().year
copyright = f"2024-{current_year}, Daily" if current_year > 2024 else "2024, Daily"
author = "Daily"
# General configuration
@@ -26,16 +28,14 @@ extensions = [
# Napoleon settings
napoleon_google_docstring = True
napoleon_numpy_docstring = False
napoleon_include_init_with_doc = True
# AutoDoc settings
autodoc_default_options = {
"members": True,
"member-order": "bysource",
"special-members": "__init__",
"undoc-members": True,
"exclude-members": "__weakref__",
"exclude-members": "__weakref__,model_config",
"no-index": True,
"show-inheritance": True,
}
@@ -145,12 +145,34 @@ autodoc_mock_imports = [
"transformers.AutoFeatureExtractor",
# Also add specific classes that are imported
"AutoFeatureExtractor",
# Sentry dependencies
"sentry_sdk",
# AWS Nova Sonic dependencies
"aws_sdk_bedrock_runtime",
"aws_sdk_bedrock_runtime.client",
"aws_sdk_bedrock_runtime.config",
"aws_sdk_bedrock_runtime.models",
"smithy_aws_core",
"smithy_aws_core.credentials_resolvers",
"smithy_aws_core.credentials_resolvers.static",
"smithy_aws_core.identity",
"smithy_core",
"smithy_core.aio",
"smithy_core.aio.eventstream",
# MCP dependencies (you may already have these)
"mcp",
"mcp.client",
"mcp.client.session_group",
"mcp.client.sse",
"mcp.client.stdio",
"mcp.ClientSession",
"mcp.StdioServerParameters",
]
# HTML output settings
html_theme = "sphinx_rtd_theme"
html_static_path = ["_static"]
autodoc_typehints = "description"
autodoc_typehints = "signature" # Show type hints in the signature only, not in the docstring
html_show_sphinx = False
@@ -249,6 +271,10 @@ def clean_title(title: str) -> str:
"playht": "PlayHT",
"xtts": "XTTS",
"lmnt": "LMNT",
"stt": "STT",
"tts": "TTS",
"llm": "LLM",
"rtvi": "RTVI",
}
# Check if the entire title is a special case

7
examples/foundational/07n-interruptible-google.py
View File

@@ -61,7 +61,12 @@ async def run_example(transport: BaseTransport, _: argparse.Namespace, handle_si
credentials=os.getenv("GOOGLE_TEST_CREDENTIALS"),
)
llm = GoogleLLMService(api_key=os.getenv("GOOGLE_API_KEY"))
llm = GoogleLLMService(
api_key=os.getenv("GOOGLE_API_KEY"),
model="gemini-2.5-flash",
# turn on thinking if you want it
# params=GoogleLLMService.InputParams(extra={"thinking_config": {"thinking_budget": 4096}}),)
)
messages = [
{

7
examples/foundational/07s-interruptible-google-audio-in.py
View File

@@ -214,7 +214,12 @@ transport_params = {
async def run_example(transport: BaseTransport, _: argparse.Namespace, handle_sigint: bool):
logger.info(f"Starting bot")
llm = GoogleLLMService(api_key=os.getenv("GOOGLE_API_KEY"), model="gemini-2.0-flash-001")
llm = GoogleLLMService(
api_key=os.getenv("GOOGLE_API_KEY"),
model="gemini-2.5-flash",
# turn on thinking if you want it
# params=GoogleLLMService.InputParams(extra={"thinking_config": {"thinking_budget": 4096}}),
)
tts = GoogleTTSService(
voice_id="en-US-Chirp3-HD-Charon",

133
examples/foundational/39c-mcp-run-http.py Normal file
View File

@@ -0,0 +1,133 @@
#
# Copyright (c) 2024–2025, Daily
#
# SPDX-License-Identifier: BSD 2-Clause License
#
import argparse
import os
from dotenv import load_dotenv
from loguru import logger
from mcp.client.session_group import StreamableHttpParameters
from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.pipeline.pipeline import Pipeline
from pipecat.pipeline.runner import PipelineRunner
from pipecat.pipeline.task import PipelineParams, PipelineTask
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContext
from pipecat.services.cartesia.tts import CartesiaTTSService
from pipecat.services.deepgram.stt import DeepgramSTTService
from pipecat.services.mcp_service import MCPClient
from pipecat.services.openai.llm import OpenAILLMService
from pipecat.transports.base_transport import BaseTransport, TransportParams
from pipecat.transports.network.fastapi_websocket import FastAPIWebsocketParams
from pipecat.transports.services.daily import DailyParams
load_dotenv(override=True)
# We store functions so objects (e.g. SileroVADAnalyzer) don't get
# instantiated. The function will be called when the desired transport gets
# selected.
transport_params = {
"daily": lambda: DailyParams(
audio_in_enabled=True,
audio_out_enabled=True,
vad_analyzer=SileroVADAnalyzer(),
),
"twilio": lambda: FastAPIWebsocketParams(
audio_in_enabled=True,
audio_out_enabled=True,
vad_analyzer=SileroVADAnalyzer(),
),
"webrtc": lambda: TransportParams(
audio_in_enabled=True,
audio_out_enabled=True,
vad_analyzer=SileroVADAnalyzer(),
),
}
async def run_example(transport: BaseTransport, _: argparse.Namespace, handle_sigint: bool):
logger.info(f"Starting bot")
stt = DeepgramSTTService(api_key=os.getenv("DEEPGRAM_API_KEY"))
tts = CartesiaTTSService(
api_key=os.getenv("CARTESIA_API_KEY"),
voice_id="71a7ad14-091c-4e8e-a314-022ece01c121", # British Reading Lady
)
llm = OpenAILLMService(api_key=os.getenv("OPENAI_API_KEY"), model="gpt-4o-mini")
try:
# Github MCP docs: https://github.com/github/github-mcp-server
# Enable Github Copilot on your GitHub account. Free tier is ok. (https://github.com/settings/copilot)
# Generate a personal access token. It must be a Fine-grained token, classic tokens are not supported. (https://github.com/settings/personal-access-tokens)
# Set permissions you want to use (eg. "all repositories", "profile: read/write", etc)
mcp = MCPClient(
server_params=StreamableHttpParameters(
url="https://api.githubcopilot.com/mcp/",
headers={"Authorization": f"Bearer {os.getenv('GITHUB_PERSONAL_ACCESS_TOKEN')}"},
)
)
except Exception as e:
logger.error(f"error setting up mcp")
logger.exception("error trace:")
tools = await mcp.register_tools(llm)
system = f"""
You are a helpful LLM in a WebRTC call.
Your goal is to answer questions about the user's GitHub repositories and account.
You have access to a number of tools provided by Github. Use any and all tools to help users.
Your output will be converted to audio so don't include special characters in your answers.
Don't overexplain what you are doing.
Just respond with short sentences when you are carrying out tool calls.
"""
messages = [{"role": "system", "content": system}]
context = OpenAILLMContext(messages, tools)
context_aggregator = llm.create_context_aggregator(context)
pipeline = Pipeline(
[
transport.input(), # Transport user input
stt,
context_aggregator.user(), # User spoken responses
llm, # LLM
tts, # TTS
transport.output(), # Transport bot output
context_aggregator.assistant(), # Assistant spoken responses and tool context
]
)
task = PipelineTask(
pipeline,
params=PipelineParams(
enable_metrics=True,
enable_usage_metrics=True,
),
)
@transport.event_handler("on_client_connected")
async def on_client_connected(transport, client):
logger.info(f"Client connected: {client}")
# Kick off the conversation.
await task.queue_frames([context_aggregator.user().get_context_frame()])
@transport.event_handler("on_client_disconnected")
async def on_client_disconnected(transport, client):
logger.info(f"Client disconnected")
await task.cancel()
runner = PipelineRunner(handle_sigint=handle_sigint)
await runner.run(task)
if __name__ == "__main__":
from pipecat.examples.run import main
main(run_example, transport_params=transport_params)

2
examples/foundational/42-interruption-config.py
View File

@@ -10,8 +10,8 @@ import os
from dotenv import load_dotenv
from loguru import logger
from pipecat.audio.interruptions.min_words_interruption_strategy import MinWordsInterruptionStrategy
from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.frames.frames import MinWordsInterruptionStrategy
from pipecat.pipeline.pipeline import Pipeline
from pipecat.pipeline.runner import PipelineRunner
from pipecat.pipeline.task import PipelineParams, PipelineTask

29
examples/freeze-test/freeze_test_bot.py
View File

@@ -7,9 +7,11 @@
import argparse
import asyncio
import os
import random
from contextlib import asynccontextmanager
from typing import Any, Dict
import sentry_sdk
import uvicorn
from dotenv import load_dotenv
from fastapi import FastAPI, Request, WebSocket
@@ -44,6 +46,7 @@ from pipecat.processors.aggregators.openai_llm_context import (
)
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.processors.frameworks.rtvi import RTVIConfig, RTVIProcessor
from pipecat.processors.metrics.sentry import SentryMetrics
from pipecat.serializers.protobuf import ProtobufFrameSerializer
from pipecat.services.cartesia.tts import CartesiaTTSService
from pipecat.services.deepgram.stt import DeepgramSTTService
@@ -125,6 +128,7 @@ class SimulateFreezeInput(FrameProcessor):
self._send_frames_task = None
async def _send_user_text(self, text: str):
self.reset_watchdog()
# Emulation as if the user has spoken and the stt transcribed
await self.push_frame(UserStartedSpeakingFrame())
await self.push_frame(StartInterruptionFrame())
@@ -149,14 +153,13 @@ class SimulateFreezeInput(FrameProcessor):
logger.debug("SimulateFreezeInput _send_frames")
await self._send_user_text("Tell me a brief history of Brazil!")
await asyncio.sleep(3)
await self._send_user_text("")
break
# i += 1
# if i >= 5:
# break
await self._send_user_text("and who has discovered it")
i += 1
if i >= 20:
break
# sleeping 1s before interrupting
# wait_time = random.uniform(1, 10)
# await asyncio.sleep(wait_time)
wait_time = random.uniform(1, 10)
await asyncio.sleep(wait_time)
except Exception as e:
logger.error(f"{self} exception receiving data: {e.__class__.__name__} ({e})")
@@ -176,6 +179,11 @@ async def run_example(websocket_client):
),
)
sentry_sdk.init(
dsn=os.getenv("SENTRY_DSN"),
traces_sample_rate=1.0,
)
freeze = SimulateFreezeInput()
stt = DeepgramSTTService(api_key=os.getenv("DEEPGRAM_API_KEY"))
@@ -183,9 +191,13 @@ async def run_example(websocket_client):
tts = CartesiaTTSService(
api_key=os.getenv("CARTESIA_API_KEY"),
voice_id="71a7ad14-091c-4e8e-a314-022ece01c121", # British Reading Lady
metrics=SentryMetrics(),
)
llm = OpenAILLMService(api_key=os.getenv("OPENAI_API_KEY"))
llm = OpenAILLMService(
api_key=os.getenv("OPENAI_API_KEY"),
metrics=SentryMetrics(),
)
rtvi = RTVIProcessor(config=RTVIConfig(config=[]))
@@ -247,6 +259,7 @@ async def run_example(websocket_client):
},
),
],
enable_watchdog_timers=True,
)
@transport.event_handler("on_client_connected")

8
pyproject.toml
View File

@@ -64,7 +64,7 @@ langchain = [ "langchain~=0.3.20", "langchain-community~=0.3.20", "langchain-ope
livekit = [ "livekit~=0.22.0", "livekit-api~=0.8.2", "tenacity~=9.0.0" ]
lmnt = [ "websockets~=13.1" ]
local = [ "pyaudio~=0.2.14" ]
mcp = [ "mcp[cli]~=1.6.0" ]
mcp = [ "mcp[cli]~=1.9.4" ]
mem0 = [ "mem0ai~=0.1.94" ]
mlx-whisper = [ "mlx-whisper~=0.4.2" ]
moondream = [ "einops~=0.8.0", "timm~=1.0.13", "transformers~=4.48.0" ]
@@ -123,9 +123,9 @@ select = [
"D", # Docstring rules
"I", # Import rules
]
# We ignore D107 because class docstrings already document __init__ parameters
# and our Sphinx configuration uses napoleon_include_init_with_doc=True
ignore = ["D107"]
[tool.ruff.lint.per-file-ignores]
"**/__init__.py" = ["D104"]
[tool.ruff.lint.pydocstyle]
convention = "google"

5
scripts/evals/run-release-evals.py
View File

@@ -111,11 +111,16 @@ TESTS_26 = [
# ("26d-gemini-multimodal-live-text.py", PROMPT_SIMPLE_MATH, None),
]
TESTS_40 = [
("40-aws-nova-sonic.py", PROMPT_SIMPLE_MATH, None),
]
TESTS = [
*TESTS_07,
*TESTS_14,
*TESTS_19,
*TESTS_26,
*TESTS_40,
]

2
src/pipecat/frames/frames.py
View File

@@ -453,8 +453,8 @@ class StartFrame(SystemFrame):
allow_interruptions: bool = False
enable_metrics: bool = False
enable_usage_metrics: bool = False
report_only_initial_ttfb: bool = False
interruption_strategies: List[BaseInterruptionStrategy] = field(default_factory=list)
report_only_initial_ttfb: bool = False
@dataclass

1
src/pipecat/metrics/metrics.py
View File

@@ -22,6 +22,7 @@ class LLMTokenUsage(BaseModel):
total_tokens: int
cache_read_input_tokens: Optional[int] = None
cache_creation_input_tokens: Optional[int] = None
reasoning_tokens: Optional[int] = None
class LLMUsageMetricsData(MetricsData):

48
src/pipecat/pipeline/parallel_pipeline.py
View File

@@ -21,6 +21,7 @@ from pipecat.frames.frames import (
from pipecat.pipeline.base_pipeline import BasePipeline
from pipecat.pipeline.pipeline import Pipeline
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor, FrameProcessorSetup
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
class ParallelPipelineSource(FrameProcessor):
@@ -76,20 +77,36 @@ class ParallelPipeline(BasePipeline):
if len(args) == 0:
raise Exception(f"ParallelPipeline needs at least one argument")
self._args = args
self._sources = []
self._sinks = []
self._pipelines = []
self._seen_ids = set()
self._endframe_counter: Dict[int, int] = {}
self._up_task = None
self._down_task = None
self._up_queue = asyncio.Queue()
self._down_queue = asyncio.Queue()
self._pipelines = []
#
# BasePipeline
#
def processors_with_metrics(self) -> List[FrameProcessor]:
return list(chain.from_iterable(p.processors_with_metrics() for p in self._pipelines))
#
# Frame processor
#
async def setup(self, setup: FrameProcessorSetup):
await super().setup(setup)
self._up_queue = WatchdogQueue(setup.task_manager)
self._down_queue = WatchdogQueue(setup.task_manager)
logger.debug(f"Creating {self} pipelines")
for processors in args:
for processors in self._args:
if not isinstance(processors, list):
raise TypeError(f"ParallelPipeline argument {processors} is not a list")
@@ -107,19 +124,6 @@ class ParallelPipeline(BasePipeline):
logger.debug(f"Finished creating {self} pipelines")
#
# BasePipeline
#
def processors_with_metrics(self) -> List[FrameProcessor]:
return list(chain.from_iterable(p.processors_with_metrics() for p in self._pipelines))
#
# Frame processor
#
async def setup(self, setup: FrameProcessorSetup):
await super().setup(setup)
await asyncio.gather(*[s.setup(setup) for s in self._sources])
await asyncio.gather(*[p.setup(setup) for p in self._pipelines])
await asyncio.gather(*[s.setup(setup) for s in self._sinks])
@@ -134,7 +138,7 @@ class ParallelPipeline(BasePipeline):
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
await self._start()
await self._start(frame)
elif isinstance(frame, EndFrame):
self._endframe_counter[frame.id] = len(self._pipelines)
elif isinstance(frame, CancelFrame):
@@ -154,7 +158,7 @@ class ParallelPipeline(BasePipeline):
elif isinstance(frame, EndFrame):
await self._stop()
async def _start(self):
async def _start(self, frame: StartFrame):
await self._create_tasks()
async def _stop(self):
@@ -202,18 +206,14 @@ class ParallelPipeline(BasePipeline):
async def _process_up_queue(self):
while True:
frame = await self._up_queue.get()
self.start_watchdog()
await self._parallel_push_frame(frame, FrameDirection.UPSTREAM)
self._up_queue.task_done()
self.reset_watchdog()
async def _process_down_queue(self):
running = True
while running:
frame = await self._down_queue.get()
self.start_watchdog()
endframe_counter = self._endframe_counter.get(frame.id, 0)
# If we have a counter, decrement it.
@@ -228,5 +228,3 @@ class ParallelPipeline(BasePipeline):
running = not (endframe_counter == 0 and isinstance(frame, EndFrame))
self._down_queue.task_done()
self.reset_watchdog()

35
src/pipecat/pipeline/sync_parallel_pipeline.py
View File

@@ -15,6 +15,7 @@ from pipecat.frames.frames import ControlFrame, EndFrame, Frame, SystemFrame
from pipecat.pipeline.base_pipeline import BasePipeline
from pipecat.pipeline.pipeline import Pipeline
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor, FrameProcessorSetup
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
@dataclass
@@ -61,15 +62,30 @@ class SyncParallelPipeline(BasePipeline):
if len(args) == 0:
raise Exception(f"SyncParallelPipeline needs at least one argument")
self._args = args
self._sinks = []
self._sources = []
self._pipelines = []
self._up_queue = asyncio.Queue()
self._down_queue = asyncio.Queue()
#
# BasePipeline
#
def processors_with_metrics(self) -> List[FrameProcessor]:
return list(chain.from_iterable(p.processors_with_metrics() for p in self._pipelines))
#
# Frame processor
#
async def setup(self, setup: FrameProcessorSetup):
await super().setup(setup)
self._up_queue = WatchdogQueue(setup.task_manager)
self._down_queue = WatchdogQueue(setup.task_manager)
logger.debug(f"Creating {self} pipelines")
for processors in args:
for processors in self._args:
if not isinstance(processors, list):
raise TypeError(f"SyncParallelPipeline argument {processors} is not a list")
@@ -92,19 +108,6 @@ class SyncParallelPipeline(BasePipeline):
logger.debug(f"Finished creating {self} pipelines")
#
# BasePipeline
#
def processors_with_metrics(self) -> List[FrameProcessor]:
return list(chain.from_iterable(p.processors_with_metrics() for p in self._pipelines))
#
# Frame processor
#
async def setup(self, setup: FrameProcessorSetup):
await super().setup(setup)
await asyncio.gather(*[s["processor"].setup(setup) for s in self._sources])
await asyncio.gather(*[p.setup(setup) for p in self._pipelines])
await asyncio.gather(*[s["processor"].setup(setup) for s in self._sinks])

71
src/pipecat/pipeline/task.py
View File

@@ -38,7 +38,13 @@ from pipecat.pipeline.base_pipeline import BasePipeline
from pipecat.pipeline.base_task import BasePipelineTask, PipelineTaskParams
from pipecat.pipeline.task_observer import TaskObserver
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor, FrameProcessorSetup
from pipecat.utils.asyncio import WATCHDOG_TIMEOUT, BaseTaskManager, TaskManager, TaskManagerParams
from pipecat.utils.asyncio.task_manager import (
WATCHDOG_TIMEOUT,
BaseTaskManager,
TaskManager,
TaskManagerParams,
)
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
from pipecat.utils.tracing.setup import is_tracing_available
from pipecat.utils.tracing.turn_trace_observer import TurnTraceObserver
@@ -188,6 +194,7 @@ class PipelineTask(BasePipelineTask):
enable_tracing: Whether to enable tracing.
enable_turn_tracking: Whether to enable turn tracking.
enable_watchdog_logging: Whether to print task processing times.
enable_watchdog_timers: Whether to enable task watchdog timers.
idle_timeout_frames: A tuple with the frames that should trigger an idle
timeout if not received withing `idle_timeout_seconds`.
idle_timeout_secs: Timeout (in seconds) to consider pipeline idle or
@@ -211,6 +218,7 @@ class PipelineTask(BasePipelineTask):
enable_tracing: bool = False,
enable_turn_tracking: bool = True,
enable_watchdog_logging: bool = False,
enable_watchdog_timers: bool = False,
idle_timeout_frames: Tuple[Type[Frame], ...] = (
BotSpeakingFrame,
LLMFullResponseEndFrame,
@@ -231,6 +239,7 @@ class PipelineTask(BasePipelineTask):
self._enable_tracing = enable_tracing and is_tracing_available()
self._enable_turn_tracking = enable_turn_tracking
self._enable_watchdog_logging = enable_watchdog_logging
self._enable_watchdog_timers = enable_watchdog_timers
self._idle_timeout_frames = idle_timeout_frames
self._idle_timeout_secs = idle_timeout_secs
self._watchdog_timeout_secs = watchdog_timeout_secs
@@ -260,19 +269,29 @@ class PipelineTask(BasePipelineTask):
self._finished = False
self._cancelled = False
# This task maneger will handle all the asyncio tasks created by this
# PipelineTask and its frame processors.
self._task_manager = task_manager or TaskManager()
# This queue receives frames coming from the pipeline upstream.
self._up_queue = asyncio.Queue()
self._up_queue = WatchdogQueue(self._task_manager)
self._process_up_task: Optional[asyncio.Task] = None
# This queue receives frames coming from the pipeline downstream.
self._down_queue = asyncio.Queue()
self._down_queue = WatchdogQueue(self._task_manager)
self._process_down_task: Optional[asyncio.Task] = None
# This queue is the queue used to push frames to the pipeline.
self._push_queue = asyncio.Queue()
self._push_queue = WatchdogQueue(self._task_manager)
self._process_push_task: Optional[asyncio.Task] = None
# This is the heartbeat queue. When a heartbeat frame is received in the
# down queue we add it to the heartbeat queue for processing.
self._heartbeat_queue = asyncio.Queue()
self._heartbeat_queue = WatchdogQueue(self._task_manager)
self._heartbeat_push_task: Optional[asyncio.Task] = None
self._heartbeat_monitor_task: Optional[asyncio.Task] = None
# This is the idle queue. When frames are received downstream they are
# put in the queue. If no frame is received the pipeline is considered
# idle.
self._idle_queue = asyncio.Queue()
self._idle_queue = WatchdogQueue(self._task_manager)
self._idle_monitor_task: Optional[asyncio.Task] = None
# This event is used to indicate a finalize frame (e.g. EndFrame,
# StopFrame) has been received in the down queue.
self._pipeline_end_event = asyncio.Event()
@@ -289,10 +308,6 @@ class PipelineTask(BasePipelineTask):
self._sink = PipelineTaskSink(self._down_queue)
pipeline.link(self._sink)
# This task maneger will handle all the asyncio tasks created by this
# PipelineTask and its frame processors.
self._task_manager = task_manager or TaskManager()
# The task observer acts as a proxy to the provided observers. This way,
# we only need to pass a single observer (using the StartFrame) which
# then just acts as a proxy.
@@ -433,7 +448,9 @@ class PipelineTask(BasePipelineTask):
# we want to cancel right away.
await self._source.push_frame(CancelFrame())
# Only cancel the push task. Everything else will be cancelled in run().
await self._task_manager.cancel_task(self._process_push_task)
if self._process_push_task:
await self._task_manager.cancel_task(self._process_push_task)
self._process_push_task = None
async def _create_tasks(self):
self._process_up_task = self._task_manager.create_task(
@@ -451,7 +468,7 @@ class PipelineTask(BasePipelineTask):
return self._process_push_task
def _maybe_start_heartbeat_tasks(self):
if self._params.enable_heartbeats:
if self._params.enable_heartbeats and self._heartbeat_push_task is None:
self._heartbeat_push_task = self._task_manager.create_task(
self._heartbeat_push_handler(), f"{self}::_heartbeat_push_handler"
)
@@ -468,20 +485,33 @@ class PipelineTask(BasePipelineTask):
async def _cancel_tasks(self):
await self._observer.stop()
await self._task_manager.cancel_task(self._process_up_task)
await self._task_manager.cancel_task(self._process_down_task)
if self._process_up_task:
await self._task_manager.cancel_task(self._process_up_task)
self._process_up_task = None
if self._process_down_task:
await self._task_manager.cancel_task(self._process_down_task)
self._process_down_task = None
await self._maybe_cancel_heartbeat_tasks()
await self._maybe_cancel_idle_task()
async def _maybe_cancel_heartbeat_tasks(self):
if self._params.enable_heartbeats:
if not self._params.enable_heartbeats:
return
if self._heartbeat_push_task:
await self._task_manager.cancel_task(self._heartbeat_push_task)
self._heartbeat_push_task = None
if self._heartbeat_monitor_task:
await self._task_manager.cancel_task(self._heartbeat_monitor_task)
self._heartbeat_monitor_task = None
async def _maybe_cancel_idle_task(self):
if self._idle_timeout_secs:
if self._idle_timeout_secs and self._idle_monitor_task:
await self._task_manager.cancel_task(self._idle_monitor_task)
self._idle_monitor_task = None
def _initial_metrics_frame(self) -> MetricsFrame:
processors = self._pipeline.processors_with_metrics()
@@ -499,6 +529,7 @@ class PipelineTask(BasePipelineTask):
mgr_params = TaskManagerParams(
loop=params.loop,
enable_watchdog_logging=self._enable_watchdog_logging,
enable_watchdog_timers=self._enable_watchdog_timers,
watchdog_timeout=self._watchdog_timeout_secs,
)
self._task_manager.setup(mgr_params)
@@ -507,6 +538,7 @@ class PipelineTask(BasePipelineTask):
clock=self._clock,
task_manager=self._task_manager,
observer=self._observer,
watchdog_timers_enabled=self._enable_watchdog_timers,
)
await self._source.setup(setup)
await self._pipeline.setup(setup)
@@ -526,8 +558,6 @@ class PipelineTask(BasePipelineTask):
await self._pipeline.cleanup()
await self._sink.cleanup()
await self._task_manager.cleanup()
async def _process_push_queue(self):
"""This is the task that runs the pipeline for the first time by sending
a StartFrame and by pushing any other frames queued by the user. It runs
@@ -536,7 +566,6 @@ class PipelineTask(BasePipelineTask):
"""
self._clock.start()
self._maybe_start_heartbeat_tasks()
self._maybe_start_idle_task()
start_frame = StartFrame(
@@ -618,6 +647,10 @@ class PipelineTask(BasePipelineTask):
if isinstance(frame, StartFrame):
await self._call_event_handler("on_pipeline_started", frame)
# Start heartbeat tasks now that StartFrame has been processed
# by all processors in the pipeline
self._maybe_start_heartbeat_tasks()
elif isinstance(frame, EndFrame):
await self._call_event_handler("on_pipeline_ended", frame)
self._pipeline_end_event.set()

8
src/pipecat/pipeline/task_observer.py
View File

@@ -11,7 +11,8 @@ from typing import Dict, List, Optional
from attr import dataclass
from pipecat.observers.base_observer import BaseObserver, FramePushed
from pipecat.utils.asyncio import BaseTaskManager
from pipecat.utils.asyncio.task_manager import BaseTaskManager
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
@dataclass
@@ -82,6 +83,9 @@ class TaskObserver(BaseObserver):
async def stop(self):
"""Stops all proxy observer tasks."""
if not self._proxies:
return
for proxy in self._proxies.values():
await self._task_manager.cancel_task(proxy.task)
@@ -93,7 +97,7 @@ class TaskObserver(BaseObserver):
return self._proxies is not None
def _create_proxy(self, observer: BaseObserver) -> Proxy:
queue = asyncio.Queue()
queue = WatchdogQueue(self._task_manager)
task = self._task_manager.create_task(
self._proxy_task_handler(queue, observer),
f"TaskObserver::{observer}::_proxy_task_handler",

27
src/pipecat/processors/aggregators/dtmf_aggregator.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""DTMF aggregation processor for converting keypad input to transcription.
This module provides a frame processor that aggregates DTMF (Dual-Tone Multi-Frequency)
keypad inputs into meaningful sequences and converts them to transcription frames
for downstream processing by LLM context aggregators.
"""
import asyncio
from typing import Optional
@@ -31,11 +38,6 @@ class DTMFAggregator(FrameProcessor):
- EndFrame or CancelFrame is received
Emits TranscriptionFrame for compatibility with existing LLM context aggregators.
Args:
timeout: Idle timeout in seconds before flushing
termination_digit: Digit that triggers immediate flush
prefix: Prefix added to DTMF sequence in transcription
"""
def __init__(
@@ -45,6 +47,14 @@ class DTMFAggregator(FrameProcessor):
prefix: str = "DTMF: ",
**kwargs,
):
"""Initialize the DTMF aggregator.
Args:
timeout: Idle timeout in seconds before flushing
termination_digit: Digit that triggers immediate flush
prefix: Prefix added to DTMF sequence in transcription
**kwargs: Additional arguments passed to FrameProcessor
"""
super().__init__(**kwargs)
self._aggregation = ""
self._idle_timeout = timeout
@@ -55,6 +65,12 @@ class DTMFAggregator(FrameProcessor):
self._aggregation_task: Optional[asyncio.Task] = None
async def process_frame(self, frame: Frame, direction: FrameDirection) -> None:
"""Process incoming frames and handle DTMF aggregation.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
@@ -119,6 +135,7 @@ class DTMFAggregator(FrameProcessor):
await asyncio.wait_for(self._digit_event.wait(), timeout=self._idle_timeout)
self._digit_event.clear()
except asyncio.TimeoutError:
self.reset_watchdog()
if self._aggregation:
await self._flush_aggregation()

26
src/pipecat/processors/aggregators/gated.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Gated frame aggregator for conditional frame accumulation.
This module provides a gated aggregator that accumulates frames based on
custom gate open/close functions, allowing for conditional frame buffering
and release in frame processing pipelines.
"""
from typing import List, Tuple
from loguru import logger
@@ -14,8 +21,11 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class GatedAggregator(FrameProcessor):
"""Accumulate frames, with custom functions to start and stop accumulation.
Yields gate-opening frame before any accumulated frames, then ensuing frames
until and not including the gate-closed frame.
until and not including the gate-closed frame. The aggregator maintains an
internal gate state that controls whether frames are passed through immediately
or accumulated for later release.
Doctest: FIXME to work with asyncio
>>> from pipecat.frames.frames import ImageRawFrame
@@ -48,6 +58,14 @@ class GatedAggregator(FrameProcessor):
start_open,
direction: FrameDirection = FrameDirection.DOWNSTREAM,
):
"""Initialize the gated aggregator.
Args:
gate_open_fn: Function that returns True when a frame should open the gate.
gate_close_fn: Function that returns True when a frame should close the gate.
start_open: Whether the gate should start in the open state.
direction: The frame direction this aggregator operates on.
"""
super().__init__()
self._gate_open_fn = gate_open_fn
self._gate_close_fn = gate_close_fn
@@ -56,6 +74,12 @@ class GatedAggregator(FrameProcessor):
self._accumulator: List[Tuple[Frame, FrameDirection]] = []
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames with gated accumulation logic.
Args:
frame: The frame to process.
direction: The direction of the frame flow.
"""
await super().process_frame(frame, direction)
# We must not block system frames.

24
src/pipecat/processors/aggregators/gated_openai_llm_context.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Gated OpenAI LLM context aggregator for controlled message flow."""
from pipecat.frames.frames import CancelFrame, EndFrame, Frame, StartFrame
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContextFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
@@ -11,12 +13,21 @@ from pipecat.sync.base_notifier import BaseNotifier
class GatedOpenAILLMContextAggregator(FrameProcessor):
"""This aggregator keeps the last received OpenAI LLM context frame and it
doesn't let it through until the notifier is notified.
"""Aggregator that gates OpenAI LLM context frames until notified.
This aggregator captures OpenAI LLM context frames and holds them until
a notifier signals that they can be released. This is useful for controlling
the flow of context frames based on external conditions or timing.
"""
def __init__(self, *, notifier: BaseNotifier, start_open: bool = False, **kwargs):
"""Initialize the gated context aggregator.
Args:
notifier: The notifier that controls when frames are released.
start_open: If True, the first context frame passes through immediately.
**kwargs: Additional arguments passed to the parent FrameProcessor.
"""
super().__init__(**kwargs)
self._notifier = notifier
self._start_open = start_open
@@ -24,6 +35,12 @@ class GatedOpenAILLMContextAggregator(FrameProcessor):
self._gate_task = None
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames, gating OpenAI LLM context frames.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
@@ -42,15 +59,18 @@ class GatedOpenAILLMContextAggregator(FrameProcessor):
await self.push_frame(frame, direction)
async def _start(self):
"""Start the gate task handler."""
if not self._gate_task:
self._gate_task = self.create_task(self._gate_task_handler())
async def _stop(self):
"""Stop the gate task handler."""
if self._gate_task:
await self.cancel_task(self._gate_task)
self._gate_task = None
async def _gate_task_handler(self):
"""Handle the gating logic by waiting for notifications and releasing frames."""
while True:
await self._notifier.wait()
if self._last_context_frame:

319
src/pipecat/processors/aggregators/llm_response.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""LLM response aggregators for handling conversation context and message aggregation.
This module provides aggregators that process and accumulate LLM responses, user inputs,
and conversation context. These aggregators handle the flow between speech-to-text,
LLM processing, and text-to-speech components in conversational AI pipelines.
"""
import asyncio
from abc import abstractmethod
from dataclasses import dataclass
@@ -54,30 +61,55 @@ from pipecat.utils.time import time_now_iso8601
@dataclass
class LLMUserAggregatorParams:
"""Parameters for configuring LLM user aggregation behavior.
Parameters:
aggregation_timeout: Maximum time in seconds to wait for additional
transcription content before pushing aggregated result. This
timeout is used only when the transcription is slow to arrive.
"""
aggregation_timeout: float = 0.5
@dataclass
class LLMAssistantAggregatorParams:
"""Parameters for configuring LLM assistant aggregation behavior.
Parameters:
expect_stripped_words: Whether to expect and handle stripped words
in text frames by adding spaces between tokens.
"""
expect_stripped_words: bool = True
class LLMFullResponseAggregator(FrameProcessor):
"""This is an LLM aggregator that aggregates a full LLM completion. It
aggregates LLM text frames (tokens) received between
`LLMFullResponseStartFrame` and `LLMFullResponseEndFrame`. Every full
completion is returned via the "on_completion" event handler:
"""Aggregates complete LLM responses between start and end frames.
@aggregator.event_handler("on_completion")
async def on_completion(
aggregator: LLMFullResponseAggregator,
completion: str,
completed: bool,
)
This aggregator collects LLM text frames (tokens) received between
`LLMFullResponseStartFrame` and `LLMFullResponseEndFrame` and provides
the complete response via an event handler.
The aggregator provides an "on_completion" event that fires when a full
completion is available:
@aggregator.event_handler("on_completion")
async def on_completion(
aggregator: LLMFullResponseAggregator,
completion: str,
completed: bool,
):
# Handle the completion
pass
"""
def __init__(self, **kwargs):
"""Initialize the LLM full response aggregator.
Args:
**kwargs: Additional arguments passed to parent FrameProcessor.
"""
super().__init__(**kwargs)
self._aggregation = ""
@@ -86,6 +118,12 @@ class LLMFullResponseAggregator(FrameProcessor):
self._register_event_handler("on_completion")
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and aggregate LLM text content.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartInterruptionFrame):
@@ -116,83 +154,123 @@ class LLMFullResponseAggregator(FrameProcessor):
class BaseLLMResponseAggregator(FrameProcessor):
"""This is the base class for all LLM response aggregators. These
aggregators process incoming frames and aggregate content until they are
ready to push the aggregation. In the case of a user, an aggregation might
be a full transcription received from the STT service.
"""Base class for all LLM response aggregators.
The LLM response aggregators also keep a store (e.g. a message list or an
LLM context) of the current conversation, that is, it stores the messages
said by the user or by the bot.
These aggregators process incoming frames and aggregate content until they are
ready to push the aggregation downstream. They maintain conversation state
and handle message flow between different components in the pipeline.
The aggregators keep a store (e.g. message list or LLM context) of the current
conversation, storing messages from both users and the bot.
"""
def __init__(self, **kwargs):
"""Initialize the base LLM response aggregator.
Args:
**kwargs: Additional arguments passed to parent FrameProcessor.
"""
super().__init__(**kwargs)
@property
@abstractmethod
def messages(self) -> List[dict]:
"""Returns the messages from the current conversation."""
"""Get the messages from the current conversation.
Returns:
List of message dictionaries representing the conversation history.
"""
pass
@property
@abstractmethod
def role(self) -> str:
"""Returns the role (e.g. user, assistant...) for this aggregator."""
"""Get the role for this aggregator.
Returns:
The role string (e.g. "user", "assistant") for this aggregator.
"""
pass
@abstractmethod
def add_messages(self, messages):
"""Add the given messages to the conversation."""
"""Add the given messages to the conversation.
Args:
messages: Messages to append to the conversation history.
"""
pass
@abstractmethod
def set_messages(self, messages):
"""Reset the conversation with the given messages."""
"""Reset the conversation with the given messages.
Args:
messages: Messages to replace the current conversation history.
"""
pass
@abstractmethod
def set_tools(self, tools):
"""Set LLM tools to be used in the current conversation."""
"""Set LLM tools to be used in the current conversation.
Args:
tools: List of tool definitions for the LLM to use.
"""
pass
@abstractmethod
def set_tool_choice(self, tool_choice):
"""Set the tool choice. This should modify the LLM context."""
"""Set the tool choice for the LLM.
Args:
tool_choice: Tool choice configuration for the LLM context.
"""
pass
@abstractmethod
async def reset(self):
"""Reset the internals of this aggregator. This should not modify the
internal messages.
"""Reset the internal state of this aggregator.
This should clear aggregation state but not modify the conversation messages.
"""
pass
@abstractmethod
async def handle_aggregation(self, aggregation: str):
"""Adds the given aggregation to the aggregator. The aggregator can use
a simple list of message or a context. It doesn't not push any frames.
"""Add the given aggregation to the conversation store.
Args:
aggregation: The aggregated text content to add to the conversation.
"""
pass
@abstractmethod
async def push_aggregation(self):
"""Pushes the current aggregation. For example, iN the case of context
aggregation this might push a new context frame.
"""Push the current aggregation downstream.
The specific frame type pushed depends on the aggregator implementation
(e.g. context frame, messages frame).
"""
pass
class LLMContextResponseAggregator(BaseLLMResponseAggregator):
"""This is a base LLM aggregator that uses an LLM context to store the
conversation. It pushes `OpenAILLMContextFrame` as an aggregation frame.
"""Base LLM aggregator that uses an OpenAI LLM context for conversation storage.
This aggregator maintains conversation state using an OpenAILLMContext and
pushes OpenAILLMContextFrame objects as aggregation frames. It provides
common functionality for context-based conversation management.
"""
def __init__(self, *, context: OpenAILLMContext, role: str, **kwargs):
"""Initialize the context response aggregator.
Args:
context: The OpenAI LLM context to use for conversation storage.
role: The role this aggregator represents (e.g. "user", "assistant").
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._context = context
self._role = role
@@ -201,46 +279,98 @@ class LLMContextResponseAggregator(BaseLLMResponseAggregator):
@property
def messages(self) -> List[dict]:
"""Get messages from the LLM context.
Returns:
List of message dictionaries from the context.
"""
return self._context.get_messages()
@property
def role(self) -> str:
"""Get the role for this aggregator.
Returns:
The role string for this aggregator.
"""
return self._role
@property
def context(self):
"""Get the OpenAI LLM context.
Returns:
The OpenAILLMContext instance used by this aggregator.
"""
return self._context
def get_context_frame(self) -> OpenAILLMContextFrame:
"""Create a context frame with the current context.
Returns:
OpenAILLMContextFrame containing the current context.
"""
return OpenAILLMContextFrame(context=self._context)
async def push_context_frame(self, direction: FrameDirection = FrameDirection.DOWNSTREAM):
"""Push a context frame in the specified direction.
Args:
direction: The direction to push the frame (upstream or downstream).
"""
frame = self.get_context_frame()
await self.push_frame(frame, direction)
def add_messages(self, messages):
"""Add messages to the context.
Args:
messages: Messages to add to the conversation context.
"""
self._context.add_messages(messages)
def set_messages(self, messages):
"""Set the context messages.
Args:
messages: Messages to replace the current context messages.
"""
self._context.set_messages(messages)
def set_tools(self, tools: List):
"""Set tools in the context.
Args:
tools: List of tool definitions to set in the context.
"""
self._context.set_tools(tools)
def set_tool_choice(self, tool_choice: Literal["none", "auto", "required"] | dict):
"""Set tool choice in the context.
Args:
tool_choice: Tool choice configuration for the context.
"""
self._context.set_tool_choice(tool_choice)
async def reset(self):
"""Reset the aggregation state."""
self._aggregation = ""
class LLMUserContextAggregator(LLMContextResponseAggregator):
"""This is a user LLM aggregator that uses an LLM context to store the
conversation. It aggregates transcriptions from the STT service and it has
logic to handle multiple scenarios where transcriptions are received between
VAD events (`UserStartedSpeakingFrame` and `UserStoppedSpeakingFrame`) or
even outside or no VAD events at all.
"""User LLM aggregator that processes speech-to-text transcriptions.
This aggregator handles the complex logic of aggregating user speech transcriptions
from STT services. It manages multiple scenarios including:
- Transcriptions received between VAD events
- Transcriptions received outside VAD events
- Interim vs final transcriptions
- User interruptions during bot speech
- Emulated VAD for whispered or short utterances
The aggregator uses timeouts to handle cases where transcriptions arrive
after VAD events or when no VAD is available.
"""
def __init__(
@@ -250,6 +380,13 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
params: Optional[LLMUserAggregatorParams] = None,
**kwargs,
):
"""Initialize the user context aggregator.
Args:
context: The OpenAI LLM context for conversation storage.
params: Configuration parameters for aggregation behavior.
**kwargs: Additional arguments. Supports deprecated 'aggregation_timeout'.
"""
super().__init__(context=context, role="user", **kwargs)
self._params = params or LLMUserAggregatorParams()
if "aggregation_timeout" in kwargs:
@@ -275,6 +412,7 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
self._aggregation_task = None
async def reset(self):
"""Reset the aggregation state and interruption strategies."""
await super().reset()
self._was_bot_speaking = False
self._seen_interim_results = False
@@ -282,9 +420,20 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
[await s.reset() for s in self._interruption_strategies]
async def handle_aggregation(self, aggregation: str):
"""Add the aggregated user text to the context.
Args:
aggregation: The aggregated user text to add as a user message.
"""
self._context.add_message({"role": self.role, "content": aggregation})
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames for user speech aggregation and context management.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
@@ -339,7 +488,7 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
await self.push_frame(frame)
async def push_aggregation(self):
"""Pushes the current aggregation based on interruption strategies and conditions."""
"""Push the current aggregation based on interruption strategies and conditions."""
if len(self._aggregation) > 0:
if self.interruption_strategies and self._bot_speaking:
should_interrupt = await self._should_interrupt_based_on_strategies()
@@ -373,7 +522,11 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
# await self.push_frame(OpenAILLMContextFrame(self._context))
async def _should_interrupt_based_on_strategies(self) -> bool:
"""Check if interruption should occur based on configured strategies."""
"""Check if interruption should occur based on configured strategies.
Returns:
True if any interruption strategy indicates interruption should occur.
"""
async def should_interrupt(strategy: BaseInterruptionStrategy):
await strategy.append_text(self._aggregation)
@@ -470,12 +623,14 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
)
self._emulating_vad = False
finally:
self.reset_watchdog()
self._aggregation_event.clear()
async def _maybe_emulate_user_speaking(self):
"""Emulate user speaking if we got a transcription but it was not
detected by VAD. Only do that if the bot is not speaking.
"""Maybe emulate user speaking based on transcription.
Emulate user speaking if we got a transcription but it was not
detected by VAD. Only do that if the bot is not speaking.
"""
# Check if we received a transcription but VAD was not able to detect
# voice (e.g. when you whisper a short utterance). In that case, we need
@@ -496,10 +651,17 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
class LLMAssistantContextAggregator(LLMContextResponseAggregator):
"""This is an assistant LLM aggregator that uses an LLM context to store the
conversation. It aggregates text frames received between
`LLMFullResponseStartFrame` and `LLMFullResponseEndFrame`.
"""Assistant LLM aggregator that processes bot responses and function calls.
This aggregator handles the complex logic of processing assistant responses including:
- Text frame aggregation between response start/end markers
- Function call lifecycle management
- Context updates with timestamps
- Tool execution and result handling
- Interruption handling during responses
The aggregator manages function calls in progress and coordinates between
text generation and tool execution phases of LLM responses.
"""
def __init__(
@@ -509,6 +671,13 @@ class LLMAssistantContextAggregator(LLMContextResponseAggregator):
params: Optional[LLMAssistantAggregatorParams] = None,
**kwargs,
):
"""Initialize the assistant context aggregator.
Args:
context: The OpenAI LLM context for conversation storage.
params: Configuration parameters for aggregation behavior.
**kwargs: Additional arguments. Supports deprecated 'expect_stripped_words'.
"""
super().__init__(context=context, role="assistant", **kwargs)
self._params = params or LLMAssistantAggregatorParams()
@@ -533,26 +702,57 @@ class LLMAssistantContextAggregator(LLMContextResponseAggregator):
"""Check if there are any function calls currently in progress.
Returns:
bool: True if function calls are in progress, False otherwise
True if function calls are in progress, False otherwise.
"""
return bool(self._function_calls_in_progress)
async def handle_aggregation(self, aggregation: str):
"""Add the aggregated assistant text to the context.
Args:
aggregation: The aggregated assistant text to add as an assistant message.
"""
self._context.add_message({"role": "assistant", "content": aggregation})
async def handle_function_call_in_progress(self, frame: FunctionCallInProgressFrame):
"""Handle a function call that is in progress.
Args:
frame: The function call in progress frame to handle.
"""
pass
async def handle_function_call_result(self, frame: FunctionCallResultFrame):
"""Handle the result of a completed function call.
Args:
frame: The function call result frame to handle.
"""
pass
async def handle_function_call_cancel(self, frame: FunctionCallCancelFrame):
"""Handle cancellation of a function call.
Args:
frame: The function call cancel frame to handle.
"""
pass
async def handle_user_image_frame(self, frame: UserImageRawFrame):
"""Handle a user image frame associated with a function call.
Args:
frame: The user image frame to handle.
"""
pass
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames for assistant response aggregation and function call management.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartInterruptionFrame):
@@ -589,6 +789,7 @@ class LLMAssistantContextAggregator(LLMContextResponseAggregator):
await self.push_frame(frame, direction)
async def push_aggregation(self):
"""Push the current assistant aggregation with timestamp."""
if not self._aggregation:
return
@@ -718,6 +919,13 @@ class LLMAssistantContextAggregator(LLMContextResponseAggregator):
class LLMUserResponseAggregator(LLMUserContextAggregator):
"""User response aggregator that outputs LLMMessagesFrame instead of context frames.
This aggregator extends LLMUserContextAggregator but pushes LLMMessagesFrame
objects downstream instead of OpenAILLMContextFrame objects. This is useful
when you need message-based output rather than context-based output.
"""
def __init__(
self,
messages: Optional[List[dict]] = None,
@@ -725,9 +933,17 @@ class LLMUserResponseAggregator(LLMUserContextAggregator):
params: Optional[LLMUserAggregatorParams] = None,
**kwargs,
):
"""Initialize the user response aggregator.
Args:
messages: Initial messages for the conversation context.
params: Configuration parameters for aggregation behavior.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(context=OpenAILLMContext(messages), params=params, **kwargs)
async def push_aggregation(self):
"""Push the aggregated user input as an LLMMessagesFrame."""
if len(self._aggregation) > 0:
await self.handle_aggregation(self._aggregation)
@@ -740,6 +956,13 @@ class LLMUserResponseAggregator(LLMUserContextAggregator):
class LLMAssistantResponseAggregator(LLMAssistantContextAggregator):
"""Assistant response aggregator that outputs LLMMessagesFrame instead of context frames.
This aggregator extends LLMAssistantContextAggregator but pushes LLMMessagesFrame
objects downstream instead of OpenAILLMContextFrame objects. This is useful
when you need message-based output rather than context-based output.
"""
def __init__(
self,
messages: Optional[List[dict]] = None,
@@ -747,9 +970,17 @@ class LLMAssistantResponseAggregator(LLMAssistantContextAggregator):
params: Optional[LLMAssistantAggregatorParams] = None,
**kwargs,
):
"""Initialize the assistant response aggregator.
Args:
messages: Initial messages for the conversation context.
params: Configuration parameters for aggregation behavior.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(context=OpenAILLMContext(messages), params=params, **kwargs)
async def push_aggregation(self):
"""Push the aggregated assistant response as an LLMMessagesFrame."""
if len(self._aggregation) > 0:
await self.handle_aggregation(self._aggregation)

165
src/pipecat/processors/aggregators/openai_llm_context.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""OpenAI LLM context management for Pipecat.
This module provides classes for managing OpenAI-specific conversation contexts,
including message handling, tool management, and image/audio processing capabilities.
"""
import base64
import copy
import io
@@ -29,7 +35,21 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class CustomEncoder(json.JSONEncoder):
"""Custom JSON encoder for handling special data types in logging.
Provides specialized encoding for io.BytesIO objects to display
readable representations in log output instead of raw binary data.
"""
def default(self, obj):
"""Encode special objects for JSON serialization.
Args:
obj: The object to encode.
Returns:
Encoded representation of the object.
"""
if isinstance(obj, io.BytesIO):
# Convert the first 8 bytes to an ASCII hex string
return f"{obj.getbuffer()[0:8].hex()}..."
@@ -37,25 +57,57 @@ class CustomEncoder(json.JSONEncoder):
class OpenAILLMContext:
"""Manages conversation context for OpenAI LLM interactions.
Handles message history, tool definitions, tool choices, and multimedia content
for OpenAI API conversations. Provides methods for message manipulation,
content formatting, and integration with various LLM adapters.
"""
def __init__(
self,
messages: Optional[List[ChatCompletionMessageParam]] = None,
tools: List[ChatCompletionToolParam] | NotGiven | ToolsSchema = NOT_GIVEN,
tool_choice: ChatCompletionToolChoiceOptionParam | NotGiven = NOT_GIVEN,
):
"""Initialize the OpenAI LLM context.
Args:
messages: Initial list of conversation messages.
tools: Available tools for the LLM to use.
tool_choice: Tool selection strategy for the LLM.
"""
self._messages: List[ChatCompletionMessageParam] = messages if messages else []
self._tool_choice: ChatCompletionToolChoiceOptionParam | NotGiven = tool_choice
self._tools: List[ChatCompletionToolParam] | NotGiven | ToolsSchema = tools
self._llm_adapter: Optional[BaseLLMAdapter] = None
def get_llm_adapter(self) -> Optional[BaseLLMAdapter]:
"""Get the current LLM adapter.
Returns:
The currently set LLM adapter, or None if not set.
"""
return self._llm_adapter
def set_llm_adapter(self, llm_adapter: BaseLLMAdapter):
"""Set the LLM adapter for context processing.
Args:
llm_adapter: The LLM adapter to use for tool conversion.
"""
self._llm_adapter = llm_adapter
@staticmethod
def from_messages(messages: List[dict]) -> "OpenAILLMContext":
"""Create a context from a list of message dictionaries.
Args:
messages: List of message dictionaries to convert to context.
Returns:
New OpenAILLMContext instance with the provided messages.
"""
context = OpenAILLMContext()
for message in messages:
@@ -66,34 +118,81 @@ class OpenAILLMContext:
@property
def messages(self) -> List[ChatCompletionMessageParam]:
"""Get the current messages list.
Returns:
List of conversation messages.
"""
return self._messages
@property
def tools(self) -> List[ChatCompletionToolParam] | NotGiven | List[Any]:
"""Get the tools list, converting through adapter if available.
Returns:
Tools list, potentially converted by the LLM adapter.
"""
if self._llm_adapter:
return self._llm_adapter.from_standard_tools(self._tools)
return self._tools
@property
def tool_choice(self) -> ChatCompletionToolChoiceOptionParam | NotGiven:
"""Get the current tool choice setting.
Returns:
The tool choice configuration.
"""
return self._tool_choice
def add_message(self, message: ChatCompletionMessageParam):
"""Add a single message to the context.
Args:
message: The message to add to the conversation history.
"""
self._messages.append(message)
def add_messages(self, messages: List[ChatCompletionMessageParam]):
"""Add multiple messages to the context.
Args:
messages: List of messages to add to the conversation history.
"""
self._messages.extend(messages)
def set_messages(self, messages: List[ChatCompletionMessageParam]):
"""Replace all messages in the context.
Args:
messages: New list of messages to replace the current history.
"""
self._messages[:] = messages
def get_messages(self) -> List[ChatCompletionMessageParam]:
"""Get a copy of the current messages list.
Returns:
List of all messages in the conversation history.
"""
return self._messages
def get_messages_json(self) -> str:
"""Get messages as a formatted JSON string.
Returns:
JSON string representation of all messages with custom encoding.
"""
return json.dumps(self._messages, cls=CustomEncoder, ensure_ascii=False, indent=2)
def get_messages_for_logging(self) -> str:
"""Get sanitized messages suitable for logging.
Removes or truncates sensitive data like image content for safe logging.
Returns:
JSON string with sanitized message content for logging.
"""
msgs = []
for message in self.messages:
msg = copy.deepcopy(message)
@@ -118,10 +217,10 @@ class OpenAILLMContext:
Since OpenAI is our standard format, this is a passthrough function.
Args:
message (dict): Message in OpenAI format
message: Message in OpenAI format.
Returns:
dict: Same message, unchanged
Same message, unchanged.
"""
return message
@@ -133,20 +232,28 @@ class OpenAILLMContext:
other LLM services that may need to return multiple messages.
Args:
obj (dict): Message in OpenAI format with either:
- Simple content: {"role": "user", "content": "Hello"}
- List content: {"role": "user", "content": [{"type": "text", "text": "Hello"}]}
obj: Message in OpenAI format with either simple string content
or structured list content.
Returns:
list: List containing the original messages, preserving whether
the content was in simple string or structured list format
List containing the original messages, preserving the content format.
"""
return [obj]
def get_messages_for_initializing_history(self):
"""Get messages for initializing conversation history.
Returns:
List of messages suitable for history initialization.
"""
return self._messages
def get_messages_for_persistent_storage(self):
"""Get messages formatted for persistent storage.
Returns:
List of messages converted to standard format for storage.
"""
messages = []
for m in self._messages:
standard_messages = self.to_standard_messages(m)
@@ -154,9 +261,19 @@ class OpenAILLMContext:
return messages
def set_tool_choice(self, tool_choice: ChatCompletionToolChoiceOptionParam | NotGiven):
"""Set the tool choice configuration.
Args:
tool_choice: Tool selection strategy for the LLM.
"""
self._tool_choice = tool_choice
def set_tools(self, tools: List[ChatCompletionToolParam] | NotGiven | ToolsSchema = NOT_GIVEN):
"""Set the available tools for the LLM.
Args:
tools: List of tools available to the LLM, or NOT_GIVEN to disable tools.
"""
if tools != NOT_GIVEN and isinstance(tools, list) and len(tools) == 0:
tools = NOT_GIVEN
self._tools = tools
@@ -164,6 +281,14 @@ class OpenAILLMContext:
def add_image_frame_message(
self, *, format: str, size: tuple[int, int], image: bytes, text: str = None
):
"""Add a message containing an image frame.
Args:
format: Image format (e.g., 'RGB', 'RGBA').
size: Image dimensions as (width, height) tuple.
image: Raw image bytes.
text: Optional text to include with the image.
"""
buffer = io.BytesIO()
Image.frombytes(format, size, image).save(buffer, format="JPEG")
encoded_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
@@ -177,10 +302,30 @@ class OpenAILLMContext:
self.add_message({"role": "user", "content": content})
def add_audio_frames_message(self, *, audio_frames: list[AudioRawFrame], text: str = None):
"""Add a message containing audio frames.
Args:
audio_frames: List of audio frame objects to include.
text: Optional text to include with the audio.
Note:
This method is currently a placeholder for future implementation.
"""
# todo: implement for OpenAI models and others
pass
def create_wav_header(self, sample_rate, num_channels, bits_per_sample, data_size):
"""Create a WAV file header for audio data.
Args:
sample_rate: Audio sample rate in Hz.
num_channels: Number of audio channels.
bits_per_sample: Bits per audio sample.
data_size: Size of audio data in bytes.
Returns:
WAV header as a bytearray.
"""
# RIFF chunk descriptor
header = bytearray()
header.extend(b"RIFF") # ChunkID
@@ -206,10 +351,14 @@ class OpenAILLMContext:
@dataclass
class OpenAILLMContextFrame(Frame):
"""Like an LLMMessagesFrame, but with extra context specific to the OpenAI
"""Frame containing OpenAI-specific LLM context.
Like an LLMMessagesFrame, but with extra context specific to the OpenAI
API. The context in this message is also mutable, and will be changed by the
OpenAIContextAggregator frame processor.
Parameters:
context: The OpenAI LLM context containing messages, tools, and configuration.
"""
context: OpenAILLMContext

25
src/pipecat/processors/aggregators/sentence.py
View File

@@ -4,17 +4,28 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Text sentence aggregation processor for Pipecat.
This module provides a frame processor that accumulates text frames into
complete sentences, only outputting when a sentence-ending pattern is detected.
"""
from pipecat.frames.frames import EndFrame, Frame, InterimTranscriptionFrame, TextFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.utils.string import match_endofsentence
class SentenceAggregator(FrameProcessor):
"""This frame processor aggregates text frames into complete sentences.
"""Aggregates text frames into complete sentences.
This processor accumulates incoming text frames until a sentence-ending
pattern is detected, then outputs the complete sentence as a single frame.
Useful for ensuring downstream processors receive coherent, complete sentences
rather than fragmented text.
Frame input/output:
TextFrame("Hello,") -> None
TextFrame(" world.") -> TextFrame("Hello world.")
TextFrame(" world.") -> TextFrame("Hello, world.")
Doctest: FIXME to work with asyncio
>>> import asyncio
@@ -29,10 +40,20 @@ class SentenceAggregator(FrameProcessor):
"""
def __init__(self):
"""Initialize the sentence aggregator.
Sets up internal state for accumulating text frames into complete sentences.
"""
super().__init__()
self._aggregation = ""
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and aggregate text into complete sentences.
Args:
frame: The incoming frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
# We ignore interim description at this point.

24
src/pipecat/processors/aggregators/user_response.py
View File

@@ -4,15 +4,39 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""User response aggregation for text frames.
This module provides an aggregator that collects user responses and outputs
them as TextFrame objects, useful for capturing and processing user input
in conversational pipelines.
"""
from pipecat.frames.frames import TextFrame
from pipecat.processors.aggregators.llm_response import LLMUserResponseAggregator
class UserResponseAggregator(LLMUserResponseAggregator):
"""Aggregates user responses into TextFrame objects.
This aggregator extends LLMUserResponseAggregator to specifically handle
user input by collecting text responses and outputting them as TextFrame
objects when the aggregation is complete.
"""
def __init__(self, **kwargs):
"""Initialize the user response aggregator.
Args:
**kwargs: Additional arguments passed to parent LLMUserResponseAggregator.
"""
super().__init__(**kwargs)
async def push_aggregation(self):
"""Push the aggregated user response as a TextFrame.
Creates a TextFrame from the current aggregation if it contains content,
resets the aggregation state, and pushes the frame downstream.
"""
if len(self._aggregation) > 0:
frame = TextFrame(self._aggregation.strip())

29
src/pipecat/processors/aggregators/vision_image_frame.py
View File

@@ -4,14 +4,22 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Vision image frame aggregation for Pipecat.
This module provides frame aggregation functionality to combine text and image
frames into vision frames for multimodal processing.
"""
from pipecat.frames.frames import Frame, InputImageRawFrame, TextFrame, VisionImageRawFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class VisionImageFrameAggregator(FrameProcessor):
"""This aggregator waits for a consecutive TextFrame and an
InputImageRawFrame. After the InputImageRawFrame arrives it will output a
VisionImageRawFrame.
"""Aggregates consecutive text and image frames into vision frames.
This aggregator waits for a consecutive TextFrame and an InputImageRawFrame.
After the InputImageRawFrame arrives it will output a VisionImageRawFrame
combining both the text and image data for multimodal processing.
>>> from pipecat.frames.frames import ImageFrame
@@ -23,14 +31,27 @@ class VisionImageFrameAggregator(FrameProcessor):
>>> asyncio.run(print_frames(aggregator, TextFrame("What do you see?")))
>>> asyncio.run(print_frames(aggregator, ImageFrame(image=bytes([]), size=(0, 0))))
VisionImageFrame, text: What do you see?, image size: 0x0, buffer size: 0 B
"""
def __init__(self):
"""Initialize the vision image frame aggregator.
The aggregator starts with no cached text, waiting for the first
TextFrame to arrive before it can create vision frames.
"""
super().__init__()
self._describe_text = None
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and aggregate text with images.
Caches TextFrames and combines them with subsequent InputImageRawFrames
to create VisionImageRawFrames. Other frames are passed through unchanged.
Args:
frame: The incoming frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, TextFrame):

28
src/pipecat/processors/async_generator.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Async generator processor for frame serialization and streaming."""
import asyncio
from typing import Any, AsyncGenerator
@@ -17,12 +19,32 @@ from pipecat.serializers.base_serializer import FrameSerializer
class AsyncGeneratorProcessor(FrameProcessor):
"""A frame processor that serializes frames and provides them via async generator.
This processor passes frames through unchanged while simultaneously serializing
them and making the serialized data available through an async generator interface.
Useful for streaming frame data to external consumers while maintaining the
normal frame processing pipeline.
"""
def __init__(self, *, serializer: FrameSerializer, **kwargs):
"""Initialize the async generator processor.
Args:
serializer: The frame serializer to use for converting frames to data.
**kwargs: Additional arguments passed to the parent FrameProcessor.
"""
super().__init__(**kwargs)
self._serializer = serializer
self._data_queue = asyncio.Queue()
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames by passing them through and queuing serialized data.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
await self.push_frame(frame, direction)
@@ -35,6 +57,12 @@ class AsyncGeneratorProcessor(FrameProcessor):
await self._data_queue.put(data)
async def generator(self) -> AsyncGenerator[Any, None]:
"""Generate serialized frame data asynchronously.
Yields:
Serialized frame data from the internal queue until a termination
signal (None) is received.
"""
running = True
while running:
data = await self._data_queue.get()

47
src/pipecat/processors/audio/audio_buffer_processor.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Audio buffer processor for managing and synchronizing audio streams.
This module provides an AudioBufferProcessor that handles buffering and synchronization
of audio from both user input and bot output sources, with support for various audio
configurations and event-driven processing.
"""
import time
from typing import Optional
@@ -37,12 +44,6 @@ class AudioBufferProcessor(FrameProcessor):
on_user_turn_audio_data: Triggered when user turn has ended, providing that user turn's audio
on_bot_turn_audio_data: Triggered when bot turn has ended, providing that bot turn's audio
Args:
sample_rate (Optional[int]): Desired output sample rate. If None, uses source rate
num_channels (int): Number of channels (1 for mono, 2 for stereo). Defaults to 1
buffer_size (int): Size of buffer before triggering events. 0 for no buffering
enable_turn_audio (bool): Whether turn audio event handlers should be triggered
Audio handling:
- Mono output (num_channels=1): User and bot audio are mixed
- Stereo output (num_channels=2): User audio on left, bot audio on right
@@ -61,6 +62,16 @@ class AudioBufferProcessor(FrameProcessor):
enable_turn_audio: bool = False,
**kwargs,
):
"""Initialize the audio buffer processor.
Args:
sample_rate: Desired output sample rate. If None, uses source rate.
num_channels: Number of channels (1 for mono, 2 for stereo). Defaults to 1.
buffer_size: Size of buffer before triggering events. 0 for no buffering.
user_continuous_stream: Deprecated parameter for backwards compatibility.
enable_turn_audio: Whether turn audio event handlers should be triggered.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._init_sample_rate = sample_rate
self._sample_rate = 0
@@ -105,7 +116,7 @@ class AudioBufferProcessor(FrameProcessor):
"""Current sample rate of the audio processor.
Returns:
int: The sample rate in Hz
The sample rate in Hz.
"""
return self._sample_rate
@@ -114,7 +125,7 @@ class AudioBufferProcessor(FrameProcessor):
"""Number of channels in the audio output.
Returns:
int: Number of channels (1 for mono, 2 for stereo)
Number of channels (1 for mono, 2 for stereo).
"""
return self._num_channels
@@ -122,7 +133,7 @@ class AudioBufferProcessor(FrameProcessor):
"""Check if both user and bot audio buffers contain data.
Returns:
bool: True if both buffers contain audio data
True if both buffers contain audio data.
"""
return self._buffer_has_audio(self._user_audio_buffer) and self._buffer_has_audio(
self._bot_audio_buffer
@@ -135,7 +146,7 @@ class AudioBufferProcessor(FrameProcessor):
on the left channel and bot audio on the right channel.
Returns:
bytes: Mixed audio data
Mixed audio data as bytes.
"""
if self._num_channels == 1:
return mix_audio(bytes(self._user_audio_buffer), bytes(self._bot_audio_buffer))
@@ -163,7 +174,12 @@ class AudioBufferProcessor(FrameProcessor):
self._recording = False
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming audio frames and manage audio buffers."""
"""Process incoming audio frames and manage audio buffers.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
# Update output sample rate if necessary.
@@ -181,10 +197,12 @@ class AudioBufferProcessor(FrameProcessor):
await self.push_frame(frame, direction)
def _update_sample_rate(self, frame: StartFrame):
"""Update the sample rate from the start frame."""
self._sample_rate = self._init_sample_rate or frame.audio_out_sample_rate
self._audio_buffer_size_1s = self._sample_rate * 2
async def _process_recording(self, frame: Frame):
"""Process audio frames for recording."""
if isinstance(frame, InputAudioRawFrame):
# Add silence if we need to.
silence = self._compute_silence(self._last_user_frame_at)
@@ -208,6 +226,7 @@ class AudioBufferProcessor(FrameProcessor):
await self._call_on_audio_data_handler()
async def _process_turn_recording(self, frame: Frame):
"""Process frames for turn-based audio recording."""
if isinstance(frame, UserStartedSpeakingFrame):
self._user_speaking = True
elif isinstance(frame, UserStoppedSpeakingFrame):
@@ -242,6 +261,7 @@ class AudioBufferProcessor(FrameProcessor):
self._bot_turn_audio_buffer += resampled
async def _call_on_audio_data_handler(self):
"""Call the audio data event handlers with buffered audio."""
if not self.has_audio() or not self._recording:
return
@@ -263,23 +283,28 @@ class AudioBufferProcessor(FrameProcessor):
self._reset_audio_buffers()
def _buffer_has_audio(self, buffer: bytearray) -> bool:
"""Check if a buffer contains audio data."""
return buffer is not None and len(buffer) > 0
def _reset_recording(self):
"""Reset recording state and buffers."""
self._reset_audio_buffers()
self._last_user_frame_at = time.time()
self._last_bot_frame_at = time.time()
def _reset_audio_buffers(self):
"""Reset all audio buffers to empty state."""
self._user_audio_buffer = bytearray()
self._bot_audio_buffer = bytearray()
self._user_turn_audio_buffer = bytearray()
self._bot_turn_audio_buffer = bytearray()
async def _resample_audio(self, frame: AudioRawFrame) -> bytes:
"""Resample audio frame to the target sample rate."""
return await self._resampler.resample(frame.audio, frame.sample_rate, self._sample_rate)
def _compute_silence(self, from_time: float) -> bytes:
"""Compute silence to insert based on time gap."""
quiet_time = time.time() - from_time
# We should get audio frames very frequently. We introduce silence only
# if there's a big enough gap of 1s.

34
src/pipecat/processors/consumer_processor.py
View File

@@ -4,20 +4,23 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Consumer processor for consuming frames from ProducerProcessor queues."""
import asyncio
from typing import Awaitable, Callable, Optional
from pipecat.frames.frames import CancelFrame, EndFrame, Frame, StartFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.processors.producer_processor import ProducerProcessor, identity_transformer
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
class ConsumerProcessor(FrameProcessor):
"""This class passes-through frames and also consumes frames from a
producer's queue. When a frame from a producer queue is received it will be
pushed to the specified direction. The frames can be transformed into a
different type of frame before being pushed.
"""Frame processor that consumes frames from a ProducerProcessor's queue.
This processor passes through frames normally while also consuming frames
from a ProducerProcessor's queue. When frames are received from the producer
queue, they are optionally transformed and pushed in the specified direction.
"""
def __init__(
@@ -28,13 +31,27 @@ class ConsumerProcessor(FrameProcessor):
direction: FrameDirection = FrameDirection.DOWNSTREAM,
**kwargs,
):
"""Initialize the consumer processor.
Args:
producer: The producer processor to consume frames from.
transformer: Function to transform frames before pushing. Defaults to identity_transformer.
direction: Direction to push consumed frames. Defaults to DOWNSTREAM.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._transformer = transformer
self._direction = direction
self._queue: asyncio.Queue = producer.add_consumer()
self._producer = producer
self._consumer_task: Optional[asyncio.Task] = None
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and handle lifecycle events.
Args:
frame: The frame to process.
direction: The direction the frame is traveling.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
@@ -47,21 +64,24 @@ class ConsumerProcessor(FrameProcessor):
await self.push_frame(frame, direction)
async def _start(self, _: StartFrame):
"""Start the consumer task and register with the producer."""
if not self._consumer_task:
self._queue: WatchdogQueue = self._producer.add_consumer()
self._consumer_task = self.create_task(self._consumer_task_handler())
async def _stop(self, _: EndFrame):
"""Stop the consumer task."""
if self._consumer_task:
await self.cancel_task(self._consumer_task)
async def _cancel(self, _: CancelFrame):
"""Cancel the consumer task."""
if self._consumer_task:
await self.cancel_task(self._consumer_task)
async def _consumer_task_handler(self):
"""Handle consuming frames from the producer queue."""
while True:
frame = await self._queue.get()
self.start_watchdog()
new_frame = await self._transformer(frame)
await self.push_frame(new_frame, self._direction)
self.reset_watchdog()

23
src/pipecat/processors/filters/frame_filter.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Frame filtering processor for the Pipecat framework."""
from typing import Tuple, Type
from pipecat.frames.frames import EndFrame, Frame, SystemFrame
@@ -11,7 +13,21 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class FrameFilter(FrameProcessor):
"""A frame processor that filters frames based on their types.
This processor acts as a selective gate in the pipeline, allowing only
frames of specified types to pass through. System and end frames are
automatically allowed to pass through to maintain pipeline integrity.
"""
def __init__(self, types: Tuple[Type[Frame], ...]):
"""Initialize the frame filter.
Args:
types: Tuple of frame types that should be allowed to pass through
the filter. All other frame types (except SystemFrame and
EndFrame) will be blocked.
"""
super().__init__()
self._types = types
@@ -20,12 +36,19 @@ class FrameFilter(FrameProcessor):
#
def _should_passthrough_frame(self, frame):
"""Determine if a frame should pass through the filter."""
if isinstance(frame, self._types):
return True
return isinstance(frame, (EndFrame, SystemFrame))
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process an incoming frame and conditionally pass it through.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if self._should_passthrough_frame(frame):

30
src/pipecat/processors/filters/function_filter.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Function-based frame filtering for Pipecat pipelines.
This module provides a processor that filters frames based on a custom function,
allowing for flexible frame filtering logic in processing pipelines.
"""
from typing import Awaitable, Callable
from pipecat.frames.frames import EndFrame, Frame, SystemFrame
@@ -11,11 +17,26 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class FunctionFilter(FrameProcessor):
"""A frame processor that filters frames using a custom function.
This processor allows frames to pass through based on the result of a
user-provided filter function. System and end frames always pass through
regardless of the filter result.
"""
def __init__(
self,
filter: Callable[[Frame], Awaitable[bool]],
direction: FrameDirection = FrameDirection.DOWNSTREAM,
):
"""Initialize the function filter.
Args:
filter: An async function that takes a Frame and returns True if the
frame should pass through, False otherwise.
direction: The direction to apply filtering. Only frames moving in
this direction will be filtered. Defaults to DOWNSTREAM.
"""
super().__init__()
self._filter = filter
self._direction = direction
@@ -27,9 +48,18 @@ class FunctionFilter(FrameProcessor):
# Ignore system frames, end frames and frames that are not following the
# direction of this gate
def _should_passthrough_frame(self, frame, direction):
"""Check if a frame should pass through without filtering."""
# Ignore system frames, end frames and frames that are not following the
# direction of this gate
return isinstance(frame, (SystemFrame, EndFrame)) or direction != self._direction
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process a frame through the filter.
Args:
frame: The frame to process.
direction: The direction the frame is moving in the pipeline.
"""
await super().process_frame(frame, direction)
passthrough = self._should_passthrough_frame(frame, direction)

19
src/pipecat/processors/filters/identity_filter.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Identity filter for transparent frame passthrough.
This module provides a simple passthrough filter that forwards all frames
without modification, useful for testing and pipeline composition.
"""
from pipecat.frames.frames import Frame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
@@ -14,10 +20,14 @@ class IdentityFilter(FrameProcessor):
This filter acts as a transparent passthrough, allowing all frames to flow
through unchanged. It can be useful when testing `ParallelPipeline` to
create pipelines that pass through frames (no frames should be repeated).
"""
def __init__(self, **kwargs):
"""Initialize the identity filter.
Args:
**kwargs: Additional arguments passed to the parent FrameProcessor.
"""
super().__init__(**kwargs)
#
@@ -25,6 +35,11 @@ class IdentityFilter(FrameProcessor):
#
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process an incoming frame by passing it through unchanged."""
"""Process an incoming frame by passing it through unchanged.
Args:
frame: The frame to process and forward.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
await self.push_frame(frame, direction)

25
src/pipecat/processors/filters/null_filter.py
View File

@@ -4,14 +4,31 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Null filter processor for blocking frame transmission.
This module provides a frame processor that blocks all frames except
system and end frames, useful for testing or temporarily stopping
frame flow in a pipeline.
"""
from pipecat.frames.frames import EndFrame, Frame, SystemFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class NullFilter(FrameProcessor):
"""This filter doesn't allow passing any frames up or downstream."""
"""A filter that blocks all frames except system and end frames.
This processor acts as a null filter, preventing frames from passing
through the pipeline while still allowing essential system and end
frames to maintain proper pipeline operation.
"""
def __init__(self, **kwargs):
"""Initialize the null filter.
Args:
**kwargs: Additional arguments passed to parent FrameProcessor.
"""
super().__init__(**kwargs)
#
@@ -19,6 +36,12 @@ class NullFilter(FrameProcessor):
#
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames, only allowing system and end frames through.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, (SystemFrame, EndFrame)):

70
src/pipecat/processors/filters/stt_mute_filter.py
View File

@@ -39,12 +39,17 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class STTMuteStrategy(Enum):
"""Strategies determining when STT should be muted.
Attributes:
FIRST_SPEECH: Mute only during first detected bot speech
MUTE_UNTIL_FIRST_BOT_COMPLETE: Start muted and remain muted until first bot speech completes
FUNCTION_CALL: Mute during function calls
ALWAYS: Mute during all bot speech
CUSTOM: Allow custom logic via callback
Each strategy defines different conditions under which speech-to-text
processing should be temporarily disabled to prevent unwanted audio
processing during specific conversation states.
Parameters:
FIRST_SPEECH: Mute STT until the first bot speech is detected.
MUTE_UNTIL_FIRST_BOT_COMPLETE: Mute STT until the first bot completes speaking,
regardless of whether it is the first speech.
FUNCTION_CALL: Mute STT during function calls to prevent interruptions.
ALWAYS: Always mute STT when the bot is speaking.
CUSTOM: Use a custom callback to determine muting logic dynamically.
"""
FIRST_SPEECH = "first_speech"
@@ -58,10 +63,15 @@ class STTMuteStrategy(Enum):
class STTMuteConfig:
"""Configuration for STT muting behavior.
Args:
strategies: Set of muting strategies to apply
Defines which muting strategies to apply and provides optional custom
callback for advanced muting logic. Multiple strategies can be combined
to create sophisticated muting behavior.
Parameters:
strategies: Set of muting strategies to apply simultaneously.
should_mute_callback: Optional callback for custom muting logic.
Only required when using STTMuteStrategy.CUSTOM
Only required when using STTMuteStrategy.CUSTOM. Called with
the STTMuteFilter instance to determine muting state.
Note:
MUTE_UNTIL_FIRST_BOT_COMPLETE and FIRST_SPEECH strategies should not be used together
@@ -69,10 +79,14 @@ class STTMuteConfig:
"""
strategies: set[STTMuteStrategy]
# Optional callback for custom muting logic
should_mute_callback: Optional[Callable[["STTMuteFilter"], Awaitable[bool]]] = None
def __post_init__(self):
"""Validate configuration after initialization.
Raises:
ValueError: If incompatible strategies are used together.
"""
if (
STTMuteStrategy.MUTE_UNTIL_FIRST_BOT_COMPLETE in self.strategies
and STTMuteStrategy.FIRST_SPEECH in self.strategies
@@ -86,15 +100,18 @@ class STTMuteFilter(FrameProcessor):
"""A processor that handles STT muting and interruption control.
This processor combines STT muting and interruption control as a coordinated
feature. When STT is muted, interruptions are automatically disabled.
Args:
config: Configuration specifying muting strategies
stt_service: STT service instance (deprecated, will be removed in future version)
**kwargs: Additional arguments passed to parent class
feature. When STT is muted, interruptions are automatically disabled by
suppressing VAD-related frames. This prevents unwanted speech detection
during bot speech, function calls, or other specified conditions.
"""
def __init__(self, *, config: STTMuteConfig, **kwargs):
"""Initialize the STT mute filter.
Args:
config: Configuration specifying muting strategies and behavior.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._config = config
self._first_speech_handled = False
@@ -104,18 +121,22 @@ class STTMuteFilter(FrameProcessor):
@property
def is_muted(self) -> bool:
"""Returns whether STT is currently muted."""
"""Check if STT is currently muted.
Returns:
True if STT is currently muted and audio frames are being suppressed.
"""
return self._is_muted
async def _handle_mute_state(self, should_mute: bool):
"""Handles both STT muting and interruption control."""
"""Handle STT muting and interruption control state changes."""
if should_mute != self.is_muted:
logger.debug(f"STTMuteFilter {'muting' if should_mute else 'unmuting'}")
self._is_muted = should_mute
await self.push_frame(STTMuteFrame(mute=should_mute))
async def _should_mute(self) -> bool:
"""Determines if STT should be muted based on current state and strategy."""
"""Determine if STT should be muted based on current state and strategies."""
for strategy in self._config.strategies:
match strategy:
case STTMuteStrategy.FUNCTION_CALL:
@@ -144,7 +165,16 @@ class STTMuteFilter(FrameProcessor):
return False
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Processes incoming frames and manages muting state."""
"""Process incoming frames and manage muting state.
Monitors conversation state through frame types and applies muting
strategies accordingly. Suppresses VAD-related frames when muted
while allowing other frames to pass through.
Args:
frame: The incoming frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
# Determine if we need to change mute state based on frame type

49
src/pipecat/processors/filters/wake_check_filter.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Wake phrase detection filter for Pipecat transcription processing.
This module provides a frame processor that filters transcription frames,
only allowing them through after wake phrases have been detected. Includes
keepalive functionality to maintain conversation flow after wake detection.
"""
import re
import time
from enum import Enum
@@ -16,23 +23,53 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class WakeCheckFilter(FrameProcessor):
"""This filter looks for wake phrases in the transcription frames and only passes through frames
after a wake phrase has been detected. It also has a keepalive timeout to allow for a brief
period of continued conversation after a wake phrase has been detected.
"""Frame processor that filters transcription frames based on wake phrase detection.
This filter monitors transcription frames for configured wake phrases and only
passes frames through after a wake phrase has been detected. Maintains a
keepalive timeout to allow continued conversation after wake detection.
"""
class WakeState(Enum):
"""Enumeration of wake detection states.
Parameters:
IDLE: No wake phrase detected, filtering active.
AWAKE: Wake phrase detected, allowing frames through.
"""
IDLE = 1
AWAKE = 2
class ParticipantState:
"""State tracking for individual participants.
Parameters:
participant_id: Unique identifier for the participant.
state: Current wake state (IDLE or AWAKE).
wake_timer: Timestamp of last wake phrase detection.
accumulator: Accumulated text for wake phrase matching.
"""
def __init__(self, participant_id: str):
"""Initialize participant state.
Args:
participant_id: Unique identifier for the participant.
"""
self.participant_id = participant_id
self.state = WakeCheckFilter.WakeState.IDLE
self.wake_timer = 0.0
self.accumulator = ""
def __init__(self, wake_phrases: List[str], keepalive_timeout: float = 3):
"""Initialize the wake phrase filter.
Args:
wake_phrases: List of wake phrases to detect in transcriptions.
keepalive_timeout: Duration in seconds to keep passing frames after
wake detection. Defaults to 3 seconds.
"""
super().__init__()
self._participant_states = {}
self._keepalive_timeout = keepalive_timeout
@@ -44,6 +81,12 @@ class WakeCheckFilter(FrameProcessor):
self._wake_patterns.append(pattern)
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames, filtering transcriptions based on wake detection.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
try:

24
src/pipecat/processors/filters/wake_notifier_filter.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Wake notifier filter for conditional frame-based notifications."""
from typing import Awaitable, Callable, Tuple, Type
from pipecat.frames.frames import Frame
@@ -12,10 +14,11 @@ from pipecat.sync.base_notifier import BaseNotifier
class WakeNotifierFilter(FrameProcessor):
"""This processor expects a list of frame types and will execute a given
callback predicate when a frame of any of those type is being processed. If
the callback returns true the notifier will be notified.
"""Frame processor that conditionally triggers notifications based on frame types and filters.
This processor monitors frames of specified types and executes a callback predicate
when such frames are processed. If the callback returns True, the associated
notifier is triggered, allowing for conditional wake-up or notification scenarios.
"""
def __init__(
@@ -26,12 +29,27 @@ class WakeNotifierFilter(FrameProcessor):
filter: Callable[[Frame], Awaitable[bool]],
**kwargs,
):
"""Initialize the wake notifier filter.
Args:
notifier: The notifier to trigger when conditions are met.
types: Tuple of frame types to monitor for potential notifications.
filter: Async callback that determines whether to trigger notification.
Should return True to trigger notification, False otherwise.
**kwargs: Additional arguments passed to parent FrameProcessor.
"""
super().__init__(**kwargs)
self._notifier = notifier
self._types = types
self._filter = filter
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames and conditionally trigger notifications.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, self._types) and await self._filter(frame):

301
src/pipecat/processors/frame_processor.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Frame processing pipeline infrastructure for Pipecat.
This module provides the core frame processing system that enables building
audio/video processing pipelines. It includes frame processors, pipeline
management, and frame flow control mechanisms.
"""
import asyncio
from dataclasses import dataclass
from enum import Enum
@@ -29,42 +36,83 @@ from pipecat.frames.frames import (
from pipecat.metrics.metrics import LLMTokenUsage, MetricsData
from pipecat.observers.base_observer import BaseObserver, FramePushed
from pipecat.processors.metrics.frame_processor_metrics import FrameProcessorMetrics
from pipecat.utils.asyncio import BaseTaskManager
from pipecat.utils.asyncio.task_manager import BaseTaskManager
from pipecat.utils.asyncio.watchdog_event import WatchdogEvent
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
from pipecat.utils.base_object import BaseObject
class FrameDirection(Enum):
"""Direction of frame flow in the processing pipeline.
Parameters:
DOWNSTREAM: Frames flowing from input to output.
UPSTREAM: Frames flowing back from output to input.
"""
DOWNSTREAM = 1
UPSTREAM = 2
@dataclass
class FrameProcessorSetup:
"""Configuration parameters for frame processor initialization.
Parameters:
clock: The clock instance for timing operations.
task_manager: The task manager for handling async operations.
observer: Optional observer for monitoring frame processing events.
watchdog_timers_enabled: Whether to enable watchdog timers by default.
"""
clock: BaseClock
task_manager: BaseTaskManager
observer: Optional[BaseObserver] = None
watchdog_timers_enabled: bool = False
class FrameProcessor(BaseObject):
"""Base class for all frame processors in the pipeline.
Frame processors are the building blocks of Pipecat pipelines. They receive
frames, process them, and pass them to the next processor in the chain.
Each processor runs in its own task and can be linked to form complex
processing pipelines.
"""
def __init__(
self,
*,
name: Optional[str] = None,
metrics: Optional[FrameProcessorMetrics] = None,
enable_watchdog_logging: Optional[bool] = None,
enable_watchdog_timers: Optional[bool] = None,
metrics: Optional[FrameProcessorMetrics] = None,
watchdog_timeout_secs: Optional[float] = None,
**kwargs,
):
"""Initialize the frame processor.
Args:
name: Optional name for this processor instance.
enable_watchdog_logging: Whether to enable watchdog logging for tasks.
enable_watchdog_timers: Whether to enable watchdog timers for tasks.
metrics: Optional metrics collector for this processor.
watchdog_timeout_secs: Timeout in seconds for watchdog operations.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(name=name)
self._parent: Optional["FrameProcessor"] = None
self._prev: Optional["FrameProcessor"] = None
self._next: Optional["FrameProcessor"] = None
# Enable watchdog timers for all tasks created by this frame processor.
self._enable_watchdog_timers = enable_watchdog_timers
# Enable watchdog logging for all tasks created by this frame processor.
self._enable_watchdog_logging = enable_watchdog_logging
# Allow this frame processor to control their tasks timeout.
self._watchdog_timeout = watchdog_timeout_secs
self._watchdog_timeout_secs = watchdog_timeout_secs
# Clock
self._clock: Optional[BaseClock] = None
@@ -101,7 +149,7 @@ class FrameProcessor(BaseObject):
# is called. To resume processing frames we need to call
# `resume_processing_frames()` which will wake up the event.
self.__should_block_frames = False
self.__input_event = asyncio.Event()
self.__input_event = None
self.__input_frame_task: Optional[asyncio.Task] = None
# Every processor in Pipecat should only output frames from a single
@@ -111,71 +159,145 @@ class FrameProcessor(BaseObject):
@property
def id(self) -> int:
"""Get the unique identifier for this processor.
Returns:
The unique integer ID of this processor.
"""
return self._id
@property
def name(self) -> str:
"""Get the name of this processor.
Returns:
The name of this processor instance.
"""
return self._name
@property
def interruptions_allowed(self):
"""Check if interruptions are allowed for this processor.
Returns:
True if interruptions are allowed.
"""
return self._allow_interruptions
@property
def metrics_enabled(self):
"""Check if metrics collection is enabled.
Returns:
True if metrics collection is enabled.
"""
return self._enable_metrics
@property
def usage_metrics_enabled(self):
"""Check if usage metrics collection is enabled.
Returns:
True if usage metrics collection is enabled.
"""
return self._enable_usage_metrics
@property
def report_only_initial_ttfb(self):
"""Check if only initial TTFB should be reported.
Returns:
True if only initial time-to-first-byte should be reported.
"""
return self._report_only_initial_ttfb
@property
def interruption_strategies(self) -> Sequence[BaseInterruptionStrategy]:
"""Get the interruption strategies for this processor.
Returns:
Sequence of interruption strategies.
"""
return self._interruption_strategies
@property
def task_manager(self) -> BaseTaskManager:
"""Get the task manager for this processor.
Returns:
The task manager instance.
Raises:
Exception: If the task manager is not initialized.
"""
if not self._task_manager:
raise Exception(f"{self} TaskManager is still not initialized.")
return self._task_manager
def can_generate_metrics(self) -> bool:
"""Check if this processor can generate metrics.
Returns:
True if this processor can generate metrics.
"""
return False
def set_core_metrics_data(self, data: MetricsData):
"""Set core metrics data for this processor.
Args:
data: The metrics data to set.
"""
self._metrics.set_core_metrics_data(data)
async def start_ttfb_metrics(self):
"""Start time-to-first-byte metrics collection."""
if self.can_generate_metrics() and self.metrics_enabled:
await self._metrics.start_ttfb_metrics(self._report_only_initial_ttfb)
async def stop_ttfb_metrics(self):
"""Stop time-to-first-byte metrics collection and push results."""
if self.can_generate_metrics() and self.metrics_enabled:
frame = await self._metrics.stop_ttfb_metrics()
if frame:
await self.push_frame(frame)
async def start_processing_metrics(self):
"""Start processing metrics collection."""
if self.can_generate_metrics() and self.metrics_enabled:
await self._metrics.start_processing_metrics()
async def stop_processing_metrics(self):
"""Stop processing metrics collection and push results."""
if self.can_generate_metrics() and self.metrics_enabled:
frame = await self._metrics.stop_processing_metrics()
if frame:
await self.push_frame(frame)
async def start_llm_usage_metrics(self, tokens: LLMTokenUsage):
"""Start LLM usage metrics collection.
Args:
tokens: Token usage information for the LLM.
"""
if self.can_generate_metrics() and self.usage_metrics_enabled:
frame = await self._metrics.start_llm_usage_metrics(tokens)
if frame:
await self.push_frame(frame)
async def start_tts_usage_metrics(self, text: str):
"""Start TTS usage metrics collection.
Args:
text: The text being processed by TTS.
"""
if self.can_generate_metrics() and self.usage_metrics_enabled:
frame = await self._metrics.start_tts_usage_metrics(text)
if frame:
await self.push_frame(frame)
async def stop_all_metrics(self):
"""Stop all active metrics collection."""
await self.stop_ttfb_metrics()
await self.stop_processing_metrics()
@@ -185,13 +307,26 @@ class FrameProcessor(BaseObject):
name: Optional[str] = None,
*,
enable_watchdog_logging: Optional[bool] = None,
enable_watchdog_timers: Optional[bool] = None,
watchdog_timeout_secs: Optional[float] = None,
) -> asyncio.Task:
"""Create a new task managed by this processor.
Args:
coroutine: The coroutine to run in the task.
name: Optional name for the task.
enable_watchdog_logging: Whether to enable watchdog logging.
enable_watchdog_timers: Whether to enable watchdog timers.
watchdog_timeout_secs: Timeout in seconds for watchdog operations.
Returns:
The created asyncio task.
"""
if name:
name = f"{self}::{name}"
else:
name = f"{self}::{coroutine.cr_code.co_name}"
return self.get_task_manager().create_task(
return self.task_manager.create_task(
coroutine,
name,
enable_watchdog_logging=(
@@ -199,31 +334,55 @@ class FrameProcessor(BaseObject):
if enable_watchdog_logging
else self._enable_watchdog_logging
),
enable_watchdog_timers=(
enable_watchdog_timers if enable_watchdog_timers else self._enable_watchdog_timers
),
watchdog_timeout=(
watchdog_timeout_secs if watchdog_timeout_secs else self._watchdog_timeout
watchdog_timeout_secs if watchdog_timeout_secs else self._watchdog_timeout_secs
),
)
async def cancel_task(self, task: asyncio.Task, timeout: Optional[float] = None):
await self.get_task_manager().cancel_task(task, timeout)
"""Cancel a task managed by this processor.
Args:
task: The task to cancel.
timeout: Optional timeout for task cancellation.
"""
await self.task_manager.cancel_task(task, timeout)
async def wait_for_task(self, task: asyncio.Task, timeout: Optional[float] = None):
await self.get_task_manager().wait_for_task(task, timeout)
"""Wait for a task to complete.
def start_watchdog(self):
self.get_task_manager().start_watchdog(asyncio.current_task())
Args:
task: The task to wait for.
timeout: Optional timeout for waiting.
"""
await self.task_manager.wait_for_task(task, timeout)
def reset_watchdog(self):
self.get_task_manager().reset_watchdog(asyncio.current_task())
"""Reset the watchdog timer for the current task."""
self.task_manager.task_reset_watchdog()
async def setup(self, setup: FrameProcessorSetup):
"""Set up the processor with required components.
Args:
setup: Configuration object containing setup parameters.
"""
self._clock = setup.clock
self._task_manager = setup.task_manager
self._observer = setup.observer
self._watchdog_timers_enabled = (
self._enable_watchdog_timers
if self._enable_watchdog_timers
else setup.watchdog_timers_enabled
)
if self._metrics is not None:
await self._metrics.setup(self._task_manager)
async def cleanup(self):
"""Clean up processor resources."""
await super().cleanup()
await self.__cancel_input_task()
await self.__cancel_push_task()
@@ -231,29 +390,52 @@ class FrameProcessor(BaseObject):
await self._metrics.cleanup()
def link(self, processor: "FrameProcessor"):
"""Link this processor to the next processor in the pipeline.
Args:
processor: The processor to link to.
"""
self._next = processor
processor._prev = self
logger.debug(f"Linking {self} -> {self._next}")
def get_event_loop(self) -> asyncio.AbstractEventLoop:
return self.get_task_manager().get_event_loop()
"""Get the event loop used by this processor.
Returns:
The asyncio event loop.
"""
return self.task_manager.get_event_loop()
def set_parent(self, parent: "FrameProcessor"):
"""Set the parent processor for this processor.
Args:
parent: The parent processor.
"""
self._parent = parent
def get_parent(self) -> Optional["FrameProcessor"]:
"""Get the parent processor.
Returns:
The parent processor, or None if no parent is set.
"""
return self._parent
def get_clock(self) -> BaseClock:
"""Get the clock used by this processor.
Returns:
The clock instance.
Raises:
Exception: If the clock is not initialized.
"""
if not self._clock:
raise Exception(f"{self} Clock is still not initialized.")
return self._clock
def get_task_manager(self) -> BaseTaskManager:
if not self._task_manager:
raise Exception(f"{self} TaskManager is still not initialized.")
return self._task_manager
async def queue_frame(
self,
frame: Frame,
@@ -262,6 +444,13 @@ class FrameProcessor(BaseObject):
Callable[["FrameProcessor", Frame, FrameDirection], Awaitable[None]]
] = None,
):
"""Queue a frame for processing.
Args:
frame: The frame to queue.
direction: The direction of frame flow.
callback: Optional callback to call after processing.
"""
# If we are cancelling we don't want to process any other frame.
if self._cancelling:
return
@@ -274,14 +463,23 @@ class FrameProcessor(BaseObject):
await self.__input_queue.put((frame, direction, callback))
async def pause_processing_frames(self):
"""Pause processing of queued frames."""
logger.trace(f"{self}: pausing frame processing")
self.__should_block_frames = True
async def resume_processing_frames(self):
"""Resume processing of queued frames."""
logger.trace(f"{self}: resuming frame processing")
self.__input_event.set()
if self.__input_event:
self.__input_event.set()
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process a frame.
Args:
frame: The frame to process.
direction: The direction of frame flow.
"""
if isinstance(frame, StartFrame):
await self.__start(frame)
elif isinstance(frame, StartInterruptionFrame):
@@ -297,9 +495,20 @@ class FrameProcessor(BaseObject):
await self.__resume(frame)
async def push_error(self, error: ErrorFrame):
"""Push an error frame upstream.
Args:
error: The error frame to push.
"""
await self.push_frame(error, FrameDirection.UPSTREAM)
async def push_frame(self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM):
"""Push a frame to the next processor in the pipeline.
Args:
frame: The frame to push.
direction: The direction to push the frame.
"""
if not self._check_started(frame):
return
@@ -309,25 +518,45 @@ class FrameProcessor(BaseObject):
await self.__push_queue.put((frame, direction))
async def __start(self, frame: StartFrame):
"""Handle the start frame to initialize processor state.
Args:
frame: The start frame containing initialization parameters.
"""
self.__started = True
self._allow_interruptions = frame.allow_interruptions
self._enable_metrics = frame.enable_metrics
self._enable_usage_metrics = frame.enable_usage_metrics
self._report_only_initial_ttfb = frame.report_only_initial_ttfb
self._interruption_strategies = frame.interruption_strategies
self._report_only_initial_ttfb = frame.report_only_initial_ttfb
self.__create_input_task()
self.__create_push_task()
async def __cancel(self, frame: CancelFrame):
"""Handle the cancel frame to stop processor operation.
Args:
frame: The cancel frame.
"""
self._cancelling = True
await self.__cancel_input_task()
await self.__cancel_push_task()
async def __pause(self, frame: FrameProcessorPauseFrame | FrameProcessorPauseUrgentFrame):
"""Handle pause frame to pause processor operation.
Args:
frame: The pause frame.
"""
if frame.processor.name == self.name:
await self.pause_processing_frames()
async def __resume(self, frame: FrameProcessorResumeFrame | FrameProcessorResumeUrgentFrame):
"""Handle resume frame to resume processor operation.
Args:
frame: The resume frame.
"""
if frame.processor.name == self.name:
await self.resume_processing_frames()
@@ -336,6 +565,7 @@ class FrameProcessor(BaseObject):
#
async def _start_interruption(self):
"""Start handling an interruption by canceling current tasks."""
try:
# Cancel the push frame task. This will stop pushing frames downstream.
await self.__cancel_push_task()
@@ -353,10 +583,17 @@ class FrameProcessor(BaseObject):
self.__create_push_task()
async def _stop_interruption(self):
"""Stop handling an interruption."""
# Nothing to do right now.
pass
async def __internal_push_frame(self, frame: Frame, direction: FrameDirection):
"""Internal method to push frames to adjacent processors.
Args:
frame: The frame to push.
direction: The direction to push the frame.
"""
try:
timestamp = self._clock.get_time() if self._clock else 0
if direction == FrameDirection.DOWNSTREAM and self._next:
@@ -389,25 +626,38 @@ class FrameProcessor(BaseObject):
await self.push_error(ErrorFrame(str(e)))
def _check_started(self, frame: Frame):
"""Check if the processor has been started.
Args:
frame: The frame being processed.
Returns:
True if the processor has been started.
"""
if not self.__started:
logger.error(f"{self} Trying to process {frame} but StartFrame not received yet")
return self.__started
def __create_input_task(self):
"""Create the input processing task."""
if not self.__input_frame_task:
self.__should_block_frames = False
if not self.__input_event:
self.__input_event = WatchdogEvent(self.task_manager)
self.__input_event.clear()
self.__input_queue = asyncio.Queue()
self.__input_queue = WatchdogQueue(self.task_manager)
self.__input_frame_task = self.create_task(self.__input_frame_task_handler())
async def __cancel_input_task(self):
"""Cancel the input processing task."""
if self.__input_frame_task:
await self.cancel_task(self.__input_frame_task)
self.__input_frame_task = None
async def __input_frame_task_handler(self):
"""Handle frames from the input queue."""
while True:
if self.__should_block_frames:
if self.__should_block_frames and self.__input_event:
logger.trace(f"{self}: frame processing paused")
await self.__input_event.wait()
self.__input_event.clear()
@@ -416,7 +666,6 @@ class FrameProcessor(BaseObject):
(frame, direction, callback) = await self.__input_queue.get()
try:
self.start_watchdog()
# Process the frame.
await self.process_frame(frame, direction)
# If this frame has an associated callback, call it now.
@@ -427,22 +676,22 @@ class FrameProcessor(BaseObject):
await self.push_error(ErrorFrame(str(e)))
finally:
self.__input_queue.task_done()
self.reset_watchdog()
def __create_push_task(self):
"""Create the frame pushing task."""
if not self.__push_frame_task:
self.__push_queue = asyncio.Queue()
self.__push_queue = WatchdogQueue(self.task_manager)
self.__push_frame_task = self.create_task(self.__push_frame_task_handler())
async def __cancel_push_task(self):
"""Cancel the frame pushing task."""
if self.__push_frame_task:
await self.cancel_task(self.__push_frame_task)
self.__push_frame_task = None
async def __push_frame_task_handler(self):
"""Handle frames from the push queue."""
while True:
(frame, direction) = await self.__push_queue.get()
self.start_watchdog()
await self.__internal_push_frame(frame, direction)
self.__push_queue.task_done()
self.reset_watchdog()

35
src/pipecat/processors/frameworks/langchain.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Langchain integration processor for Pipecat."""
from typing import Optional, Union
from loguru import logger
@@ -26,16 +28,40 @@ except ModuleNotFoundError as e:
class LangchainProcessor(FrameProcessor):
"""Processor that integrates Langchain runnables with Pipecat's frame pipeline.
This processor takes LLM message frames, extracts the latest user message,
and processes it through a Langchain runnable chain. The response is streamed
back as text frames with appropriate response markers.
"""
def __init__(self, chain: Runnable, transcript_key: str = "input"):
"""Initialize the Langchain processor.
Args:
chain: The Langchain runnable to use for processing messages.
transcript_key: The key to use when passing input to the chain.
"""
super().__init__()
self._chain = chain
self._transcript_key = transcript_key
self._participant_id: Optional[str] = None
def set_participant_id(self, participant_id: str):
"""Set the participant ID for session tracking.
Args:
participant_id: The participant ID to use for session configuration.
"""
self._participant_id = participant_id
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and handle LLM message frames.
Args:
frame: The incoming frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, LLMMessagesFrame):
@@ -50,6 +76,14 @@ class LangchainProcessor(FrameProcessor):
@staticmethod
def __get_token_value(text: Union[str, AIMessageChunk]) -> str:
"""Extract token value from various text types.
Args:
text: The text or message chunk to extract value from.
Returns:
The extracted string value.
"""
match text:
case str():
return text
@@ -59,6 +93,7 @@ class LangchainProcessor(FrameProcessor):
return ""
async def _ainvoke(self, text: str):
"""Invoke the Langchain runnable with the provided text."""
logger.debug(f"Invoking chain with {text}")
await self.push_frame(LLMFullResponseStartFrame())
try:

391
src/pipecat/processors/frameworks/rtvi.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""RTVI (Real-Time Voice Interface) protocol implementation for Pipecat.
This module provides the RTVI protocol implementation for real-time voice interactions
between clients and AI agents. It includes message handling, action processing,
and frame observation for the RTVI protocol.
"""
import asyncio
import base64
from dataclasses import dataclass
@@ -67,6 +74,7 @@ from pipecat.services.llm_service import (
from pipecat.transports.base_input import BaseInputTransport
from pipecat.transports.base_output import BaseOutputTransport
from pipecat.transports.base_transport import BaseTransport
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
from pipecat.utils.string import match_endofsentence
RTVI_PROTOCOL_VERSION = "0.3.0"
@@ -78,6 +86,12 @@ ActionResult = Union[bool, int, float, str, list, dict]
class RTVIServiceOption(BaseModel):
"""Configuration option for an RTVI service.
Defines a configurable option that can be set for an RTVI service,
including its name, type, and handler function.
"""
name: str
type: Literal["bool", "number", "string", "array", "object"]
handler: Callable[["RTVIProcessor", str, "RTVIServiceOptionConfig"], Awaitable[None]] = Field(
@@ -86,11 +100,18 @@ class RTVIServiceOption(BaseModel):
class RTVIService(BaseModel):
"""An RTVI service definition.
Represents a service that can be configured and used within the RTVI protocol,
containing a name and list of configurable options.
"""
name: str
options: List[RTVIServiceOption]
_options_dict: Dict[str, RTVIServiceOption] = PrivateAttr(default={})
def model_post_init(self, __context: Any) -> None:
"""Initialize the options dictionary after model creation."""
self._options_dict = {}
for option in self.options:
self._options_dict[option.name] = option
@@ -98,16 +119,32 @@ class RTVIService(BaseModel):
class RTVIActionArgumentData(BaseModel):
"""Data for an RTVI action argument.
Contains the name and value of an argument passed to an RTVI action.
"""
name: str
value: Any
class RTVIActionArgument(BaseModel):
"""Definition of an RTVI action argument.
Specifies the name and expected type of an argument for an RTVI action.
"""
name: str
type: Literal["bool", "number", "string", "array", "object"]
class RTVIAction(BaseModel):
"""An RTVI action definition.
Represents an action that can be executed within the RTVI protocol,
including its service, name, arguments, and handler function.
"""
service: str
action: str
arguments: List[RTVIActionArgument] = Field(default_factory=list)
@@ -118,6 +155,7 @@ class RTVIAction(BaseModel):
_arguments_dict: Dict[str, RTVIActionArgument] = PrivateAttr(default={})
def model_post_init(self, __context: Any) -> None:
"""Initialize the arguments dictionary after model creation."""
self._arguments_dict = {}
for arg in self.arguments:
self._arguments_dict[arg.name] = arg
@@ -125,16 +163,31 @@ class RTVIAction(BaseModel):
class RTVIServiceOptionConfig(BaseModel):
"""Configuration value for an RTVI service option.
Contains the name and value to set for a specific service option.
"""
name: str
value: Any
class RTVIServiceConfig(BaseModel):
"""Configuration for an RTVI service.
Contains the service name and list of option configurations to apply.
"""
service: str
options: List[RTVIServiceOptionConfig]
class RTVIConfig(BaseModel):
"""Complete RTVI configuration.
Contains the full configuration for all RTVI services.
"""
config: List[RTVIServiceConfig]
@@ -144,16 +197,31 @@ class RTVIConfig(BaseModel):
class RTVIUpdateConfig(BaseModel):
"""Request to update RTVI configuration.
Contains new configuration settings and whether to interrupt the bot.
"""
config: List[RTVIServiceConfig]
interrupt: bool = False
class RTVIActionRunArgument(BaseModel):
"""Argument for running an RTVI action.
Contains the name and value of an argument to pass to an action.
"""
name: str
value: Any
class RTVIActionRun(BaseModel):
"""Request to run an RTVI action.
Contains the service, action name, and optional arguments.
"""
service: str
action: str
arguments: Optional[List[RTVIActionRunArgument]] = None
@@ -161,11 +229,23 @@ class RTVIActionRun(BaseModel):
@dataclass
class RTVIActionFrame(DataFrame):
"""Frame containing an RTVI action to execute.
Parameters:
rtvi_action_run: The action to execute.
message_id: Optional message ID for response correlation.
"""
rtvi_action_run: RTVIActionRun
message_id: Optional[str] = None
class RTVIMessage(BaseModel):
"""Base RTVI message structure.
Represents the standard format for RTVI protocol messages.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: str
id: str
@@ -178,10 +258,20 @@ class RTVIMessage(BaseModel):
class RTVIErrorResponseData(BaseModel):
"""Data for an RTVI error response.
Contains the error message to send back to the client.
"""
error: str
class RTVIErrorResponse(BaseModel):
"""RTVI error response message.
Sent in response to a client request that resulted in an error.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["error-response"] = "error-response"
id: str
@@ -189,21 +279,41 @@ class RTVIErrorResponse(BaseModel):
class RTVIErrorData(BaseModel):
"""Data for an RTVI error event.
Contains error information including whether it's fatal.
"""
error: str
fatal: bool
class RTVIError(BaseModel):
"""RTVI error event message.
Sent when an error occurs that isn't in response to a specific request.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["error"] = "error"
data: RTVIErrorData
class RTVIDescribeConfigData(BaseModel):
"""Data for describing available RTVI configuration.
Contains the list of available services and their options.
"""
config: List[RTVIService]
class RTVIDescribeConfig(BaseModel):
"""Message describing available RTVI configuration.
Sent in response to a describe-config request.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["config-available"] = "config-available"
id: str
@@ -211,10 +321,20 @@ class RTVIDescribeConfig(BaseModel):
class RTVIDescribeActionsData(BaseModel):
"""Data for describing available RTVI actions.
Contains the list of available actions that can be executed.
"""
actions: List[RTVIAction]
class RTVIDescribeActions(BaseModel):
"""Message describing available RTVI actions.
Sent in response to a describe-actions request.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["actions-available"] = "actions-available"
id: str
@@ -222,6 +342,11 @@ class RTVIDescribeActions(BaseModel):
class RTVIConfigResponse(BaseModel):
"""Response containing current RTVI configuration.
Sent in response to a get-config request.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["config"] = "config"
id: str
@@ -229,10 +354,20 @@ class RTVIConfigResponse(BaseModel):
class RTVIActionResponseData(BaseModel):
"""Data for an RTVI action response.
Contains the result of executing an action.
"""
result: ActionResult
class RTVIActionResponse(BaseModel):
"""Response to an RTVI action execution.
Sent after successfully executing an action.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["action-response"] = "action-response"
id: str
@@ -240,11 +375,21 @@ class RTVIActionResponse(BaseModel):
class RTVIBotReadyData(BaseModel):
"""Data for bot ready notification.
Contains protocol version and initial configuration.
"""
version: str
config: List[RTVIServiceConfig]
class RTVIBotReady(BaseModel):
"""Message indicating bot is ready for interaction.
Sent after bot initialization is complete.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-ready"] = "bot-ready"
id: str
@@ -252,28 +397,53 @@ class RTVIBotReady(BaseModel):
class RTVILLMFunctionCallMessageData(BaseModel):
"""Data for LLM function call notification.
Contains function call details including name, ID, and arguments.
"""
function_name: str
tool_call_id: str
args: Mapping[str, Any]
class RTVILLMFunctionCallMessage(BaseModel):
"""Message notifying of an LLM function call.
Sent when the LLM makes a function call.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["llm-function-call"] = "llm-function-call"
data: RTVILLMFunctionCallMessageData
class RTVILLMFunctionCallStartMessageData(BaseModel):
"""Data for LLM function call start notification.
Contains the function name being called.
"""
function_name: str
class RTVILLMFunctionCallStartMessage(BaseModel):
"""Message notifying that an LLM function call has started.
Sent when the LLM begins a function call.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["llm-function-call-start"] = "llm-function-call-start"
data: RTVILLMFunctionCallStartMessageData
class RTVILLMFunctionCallResultData(BaseModel):
"""Data for LLM function call result.
Contains function call details and result.
"""
function_name: str
tool_call_id: str
arguments: dict
@@ -281,60 +451,103 @@ class RTVILLMFunctionCallResultData(BaseModel):
class RTVIBotLLMStartedMessage(BaseModel):
"""Message indicating bot LLM processing has started."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-llm-started"] = "bot-llm-started"
class RTVIBotLLMStoppedMessage(BaseModel):
"""Message indicating bot LLM processing has stopped."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-llm-stopped"] = "bot-llm-stopped"
class RTVIBotTTSStartedMessage(BaseModel):
"""Message indicating bot TTS processing has started."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-tts-started"] = "bot-tts-started"
class RTVIBotTTSStoppedMessage(BaseModel):
"""Message indicating bot TTS processing has stopped."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-tts-stopped"] = "bot-tts-stopped"
class RTVITextMessageData(BaseModel):
"""Data for text-based RTVI messages.
Contains text content.
"""
text: str
class RTVIBotTranscriptionMessage(BaseModel):
"""Message containing bot transcription text.
Sent when the bot's speech is transcribed.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-transcription"] = "bot-transcription"
data: RTVITextMessageData
class RTVIBotLLMTextMessage(BaseModel):
"""Message containing bot LLM text output.
Sent when the bot's LLM generates text.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-llm-text"] = "bot-llm-text"
data: RTVITextMessageData
class RTVIBotTTSTextMessage(BaseModel):
"""Message containing bot TTS text output.
Sent when text is being processed by TTS.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-tts-text"] = "bot-tts-text"
data: RTVITextMessageData
class RTVIAudioMessageData(BaseModel):
"""Data for audio-based RTVI messages.
Contains audio data and metadata.
"""
audio: str
sample_rate: int
num_channels: int
class RTVIBotTTSAudioMessage(BaseModel):
"""Message containing bot TTS audio output.
Sent when the bot's TTS generates audio.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-tts-audio"] = "bot-tts-audio"
data: RTVIAudioMessageData
class RTVIUserTranscriptionMessageData(BaseModel):
"""Data for user transcription messages.
Contains transcription text and metadata.
"""
text: str
user_id: str
timestamp: str
@@ -342,44 +555,72 @@ class RTVIUserTranscriptionMessageData(BaseModel):
class RTVIUserTranscriptionMessage(BaseModel):
"""Message containing user transcription.
Sent when user speech is transcribed.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["user-transcription"] = "user-transcription"
data: RTVIUserTranscriptionMessageData
class RTVIUserLLMTextMessage(BaseModel):
"""Message containing user text input for LLM.
Sent when user text is processed by the LLM.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["user-llm-text"] = "user-llm-text"
data: RTVITextMessageData
class RTVIUserStartedSpeakingMessage(BaseModel):
"""Message indicating user has started speaking."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["user-started-speaking"] = "user-started-speaking"
class RTVIUserStoppedSpeakingMessage(BaseModel):
"""Message indicating user has stopped speaking."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["user-stopped-speaking"] = "user-stopped-speaking"
class RTVIBotStartedSpeakingMessage(BaseModel):
"""Message indicating bot has started speaking."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-started-speaking"] = "bot-started-speaking"
class RTVIBotStoppedSpeakingMessage(BaseModel):
"""Message indicating bot has stopped speaking."""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["bot-stopped-speaking"] = "bot-stopped-speaking"
class RTVIMetricsMessage(BaseModel):
"""Message containing performance metrics.
Sent to provide performance and usage metrics.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["metrics"] = "metrics"
data: Mapping[str, Any]
class RTVIServerMessage(BaseModel):
"""Generic server message.
Used for custom server-to-client messages.
"""
label: RTVIMessageLiteral = RTVI_MESSAGE_LABEL
type: Literal["server-message"] = "server-message"
data: Any
@@ -387,28 +628,32 @@ class RTVIServerMessage(BaseModel):
@dataclass
class RTVIServerMessageFrame(SystemFrame):
"""A frame for sending server messages to the client."""
"""A frame for sending server messages to the client.
Parameters:
data: The message data to send to the client.
"""
data: Any
def __str__(self):
"""String representation of the RTVI server message frame."""
return f"{self.name}(data: {self.data})"
@dataclass
class RTVIObserverParams:
"""
Parameters for configuring RTVI Observer behavior.
"""Parameters for configuring RTVI Observer behavior.
Attributes:
bot_llm_enabled (bool): Indicates if the bot's LLM messages should be sent.
bot_tts_enabled (bool): Indicates if the bot's TTS messages should be sent.
bot_speaking_enabled (bool): Indicates if the bot's started/stopped speaking messages should be sent.
user_llm_enabled (bool): Indicates if the user's LLM input messages should be sent.
user_speaking_enabled (bool): Indicates if the user's started/stopped speaking messages should be sent.
user_transcription_enabled (bool): Indicates if user's transcription messages should be sent.
metrics_enabled (bool): Indicates if metrics messages should be sent.
errors_enabled (bool): Indicates if errors messages should be sent.
Parameters:
bot_llm_enabled: Indicates if the bot's LLM messages should be sent.
bot_tts_enabled: Indicates if the bot's TTS messages should be sent.
bot_speaking_enabled: Indicates if the bot's started/stopped speaking messages should be sent.
user_llm_enabled: Indicates if the user's LLM input messages should be sent.
user_speaking_enabled: Indicates if the user's started/stopped speaking messages should be sent.
user_transcription_enabled: Indicates if user's transcription messages should be sent.
metrics_enabled: Indicates if metrics messages should be sent.
errors_enabled: Indicates if errors messages should be sent.
"""
bot_llm_enabled: bool = True
@@ -431,15 +676,18 @@ class RTVIObserver(BaseObserver):
Note:
This observer only handles outgoing messages. Incoming RTVI client messages
are handled by the RTVIProcessor.
Args:
rtvi (RTVIProcessor): The RTVI processor to push frames to.
params (RTVIObserverParams): Settings to enable/disable specific messages.
"""
def __init__(
self, rtvi: "RTVIProcessor", *, params: Optional[RTVIObserverParams] = None, **kwargs
):
"""Initialize the RTVI observer.
Args:
rtvi: The RTVI processor to push frames to.
params: Settings to enable/disable specific messages.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._rtvi = rtvi
self._params = params or RTVIObserverParams()
@@ -451,11 +699,7 @@ class RTVIObserver(BaseObserver):
"""Process a frame being pushed through the pipeline.
Args:
src: Source processor pushing the frame
dst: Destination processor receiving the frame
frame: The frame being pushed
direction: Direction of frame flow in pipeline
timestamp: Time when frame was pushed
data: Frame push event data containing source, frame, direction, and timestamp.
"""
src = data.source
frame = data.frame
@@ -516,13 +760,14 @@ class RTVIObserver(BaseObserver):
"""Push an urgent transport message to the RTVI processor.
Args:
model: The message model to send
exclude_none: Whether to exclude None values from the model dump
model: The message model to send.
exclude_none: Whether to exclude None values from the model dump.
"""
frame = TransportMessageUrgentFrame(message=model.model_dump(exclude_none=exclude_none))
await self._rtvi.push_frame(frame)
async def _push_bot_transcription(self):
"""Push accumulated bot transcription as a message."""
if len(self._bot_transcription) > 0:
message = RTVIBotTranscriptionMessage(
data=RTVITextMessageData(text=self._bot_transcription)
@@ -531,6 +776,7 @@ class RTVIObserver(BaseObserver):
self._bot_transcription = ""
async def _handle_interruptions(self, frame: Frame):
"""Handle user speaking interruption frames."""
message = None
if isinstance(frame, UserStartedSpeakingFrame):
message = RTVIUserStartedSpeakingMessage()
@@ -541,6 +787,7 @@ class RTVIObserver(BaseObserver):
await self.push_transport_message_urgent(message)
async def _handle_bot_speaking(self, frame: Frame):
"""Handle bot speaking event frames."""
message = None
if isinstance(frame, BotStartedSpeakingFrame):
message = RTVIBotStartedSpeakingMessage()
@@ -551,6 +798,7 @@ class RTVIObserver(BaseObserver):
await self.push_transport_message_urgent(message)
async def _handle_llm_text_frame(self, frame: LLMTextFrame):
"""Handle LLM text output frames."""
message = RTVIBotLLMTextMessage(data=RTVITextMessageData(text=frame.text))
await self.push_transport_message_urgent(message)
@@ -559,6 +807,7 @@ class RTVIObserver(BaseObserver):
await self._push_bot_transcription()
async def _handle_user_transcriptions(self, frame: Frame):
"""Handle user transcription frames."""
message = None
if isinstance(frame, TranscriptionFrame):
message = RTVIUserTranscriptionMessage(
@@ -607,6 +856,7 @@ class RTVIObserver(BaseObserver):
logger.warning(f"Caught an error while trying to handle context: {e}")
async def _handle_metrics(self, frame: MetricsFrame):
"""Handle metrics frames and convert to RTVI metrics messages."""
metrics = {}
for d in frame.data:
if isinstance(d, TTFBMetricsData):
@@ -631,6 +881,13 @@ class RTVIObserver(BaseObserver):
class RTVIProcessor(FrameProcessor):
"""Main processor for handling RTVI protocol messages and actions.
This processor manages the RTVI protocol communication including client-server
handshaking, configuration management, action execution, and message routing.
It serves as the central hub for RTVI protocol operations.
"""
def __init__(
self,
*,
@@ -638,6 +895,13 @@ class RTVIProcessor(FrameProcessor):
transport: Optional[BaseTransport] = None,
**kwargs,
):
"""Initialize the RTVI processor.
Args:
config: Initial RTVI configuration.
transport: Transport layer for communication.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._config = config or RTVIConfig(config=[])
@@ -650,11 +914,9 @@ class RTVIProcessor(FrameProcessor):
self._registered_services: Dict[str, RTVIService] = {}
# A task to process incoming action frames.
self._action_queue = asyncio.Queue()
self._action_task: Optional[asyncio.Task] = None
# A task to process incoming transport messages.
self._message_queue = asyncio.Queue()
self._message_task: Optional[asyncio.Task] = None
self._register_event_handler("on_bot_started")
@@ -669,34 +931,67 @@ class RTVIProcessor(FrameProcessor):
self._input_transport.enable_audio_in_stream_on_start(False)
def register_action(self, action: RTVIAction):
"""Register an action that can be executed via RTVI.
Args:
action: The action to register.
"""
id = self._action_id(action.service, action.action)
self._registered_actions[id] = action
def register_service(self, service: RTVIService):
"""Register a service that can be configured via RTVI.
Args:
service: The service to register.
"""
self._registered_services[service.name] = service
async def set_client_ready(self):
"""Mark the client as ready and trigger the ready event."""
self._client_ready = True
await self._call_event_handler("on_client_ready")
async def set_bot_ready(self):
"""Mark the bot as ready and send the bot-ready message."""
self._bot_ready = True
await self._update_config(self._config, False)
await self._send_bot_ready()
def set_errors_enabled(self, enabled: bool):
"""Enable or disable error message sending.
Args:
enabled: Whether to send error messages.
"""
self._errors_enabled = enabled
async def interrupt_bot(self):
"""Send a bot interruption frame upstream."""
await self.push_frame(BotInterruptionFrame(), FrameDirection.UPSTREAM)
async def send_error(self, error: str):
"""Send an error message to the client.
Args:
error: The error message to send.
"""
await self._send_error_frame(ErrorFrame(error=error))
async def handle_message(self, message: RTVIMessage):
"""Handle an incoming RTVI message.
Args:
message: The RTVI message to handle.
"""
await self._message_queue.put(message)
async def handle_function_call(self, params: FunctionCallParams):
"""Handle a function call from the LLM.
Args:
params: The function call parameters.
"""
fn = RTVILLMFunctionCallMessageData(
function_name=params.function_name,
tool_call_id=params.tool_call_id,
@@ -708,6 +1003,16 @@ class RTVIProcessor(FrameProcessor):
async def handle_function_call_start(
self, function_name: str, llm: FrameProcessor, context: OpenAILLMContext
):
"""Handle the start of a function call from the LLM.
Args:
function_name: Name of the function being called.
llm: The LLM processor making the call.
context: The LLM context.
Note:
This method is deprecated. Use handle_function_call() instead.
"""
import warnings
with warnings.catch_warnings():
@@ -722,6 +1027,12 @@ class RTVIProcessor(FrameProcessor):
await self._push_transport_message(message, exclude_none=False)
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames through the RTVI processor.
Args:
frame: The frame to process.
direction: The direction of frame flow.
"""
await super().process_frame(frame, direction)
# Specific system frames
@@ -755,19 +1066,25 @@ class RTVIProcessor(FrameProcessor):
await self.push_frame(frame, direction)
async def _start(self, frame: StartFrame):
"""Start the RTVI processor tasks."""
if not self._action_task:
self._action_queue = WatchdogQueue(self.task_manager)
self._action_task = self.create_task(self._action_task_handler())
if not self._message_task:
self._message_queue = WatchdogQueue(self.task_manager)
self._message_task = self.create_task(self._message_task_handler())
await self._call_event_handler("on_bot_started")
async def _stop(self, frame: EndFrame):
"""Stop the RTVI processor tasks."""
await self._cancel_tasks()
async def _cancel(self, frame: CancelFrame):
"""Cancel the RTVI processor tasks."""
await self._cancel_tasks()
async def _cancel_tasks(self):
"""Cancel all running tasks."""
if self._action_task:
await self.cancel_task(self._action_task)
self._action_task = None
@@ -777,26 +1094,26 @@ class RTVIProcessor(FrameProcessor):
self._message_task = None
async def _push_transport_message(self, model: BaseModel, exclude_none: bool = True):
"""Push a transport message frame."""
frame = TransportMessageUrgentFrame(message=model.model_dump(exclude_none=exclude_none))
await self.push_frame(frame)
async def _action_task_handler(self):
"""Handle incoming action frames."""
while True:
frame = await self._action_queue.get()
self.start_watchdog()
await self._handle_action(frame.message_id, frame.rtvi_action_run)
self._action_queue.task_done()
self.reset_watchdog()
async def _message_task_handler(self):
"""Handle incoming transport messages."""
while True:
message = await self._message_queue.get()
self.start_watchdog()
await self._handle_message(message)
self._message_queue.task_done()
self.reset_watchdog()
async def _handle_transport_message(self, frame: TransportMessageUrgentFrame):
"""Handle an incoming transport message frame."""
try:
transport_message = frame.message
if transport_message.get("label") != RTVI_MESSAGE_LABEL:
@@ -809,6 +1126,7 @@ class RTVIProcessor(FrameProcessor):
logger.warning(f"Invalid RTVI transport message: {e}")
async def _handle_message(self, message: RTVIMessage):
"""Handle a parsed RTVI message."""
try:
match message.type:
case "client-ready":
@@ -845,6 +1163,7 @@ class RTVIProcessor(FrameProcessor):
logger.warning(f"Exception processing message: {e}")
async def _handle_client_ready(self, request_id: str):
"""Handle a client-ready message."""
logger.debug("Received client-ready")
if self._input_transport:
await self._input_transport.start_audio_in_streaming()
@@ -853,6 +1172,7 @@ class RTVIProcessor(FrameProcessor):
await self.set_client_ready()
async def _handle_audio_buffer(self, data):
"""Handle incoming audio buffer data."""
if not self._input_transport:
return
@@ -874,20 +1194,24 @@ class RTVIProcessor(FrameProcessor):
logger.error(f"Error processing audio buffer: {e}")
async def _handle_describe_config(self, request_id: str):
"""Handle a describe-config request."""
services = list(self._registered_services.values())
message = RTVIDescribeConfig(id=request_id, data=RTVIDescribeConfigData(config=services))
await self._push_transport_message(message)
async def _handle_describe_actions(self, request_id: str):
"""Handle a describe-actions request."""
actions = list(self._registered_actions.values())
message = RTVIDescribeActions(id=request_id, data=RTVIDescribeActionsData(actions=actions))
await self._push_transport_message(message)
async def _handle_get_config(self, request_id: str):
"""Handle a get-config request."""
message = RTVIConfigResponse(id=request_id, data=self._config)
await self._push_transport_message(message)
def _update_config_option(self, service: str, config: RTVIServiceOptionConfig):
"""Update a specific configuration option."""
for service_config in self._config.config:
if service_config.service == service:
for option_config in service_config.options:
@@ -899,6 +1223,7 @@ class RTVIProcessor(FrameProcessor):
service_config.options.append(config)
async def _update_service_config(self, config: RTVIServiceConfig):
"""Update configuration for a specific service."""
service = self._registered_services[config.service]
for option in config.options:
handler = service._options_dict[option.name].handler
@@ -906,16 +1231,19 @@ class RTVIProcessor(FrameProcessor):
self._update_config_option(service.name, option)
async def _update_config(self, data: RTVIConfig, interrupt: bool):
"""Update the RTVI configuration."""
if interrupt:
await self.interrupt_bot()
for service_config in data.config:
await self._update_service_config(service_config)
async def _handle_update_config(self, request_id: str, data: RTVIUpdateConfig):
"""Handle an update-config request."""
await self._update_config(RTVIConfig(config=data.config), data.interrupt)
await self._handle_get_config(request_id)
async def _handle_function_call_result(self, data):
"""Handle a function call result from the client."""
frame = FunctionCallResultFrame(
function_name=data.function_name,
tool_call_id=data.tool_call_id,
@@ -925,6 +1253,7 @@ class RTVIProcessor(FrameProcessor):
await self.push_frame(frame)
async def _handle_action(self, request_id: Optional[str], data: RTVIActionRun):
"""Handle an action execution request."""
action_id = self._action_id(data.service, data.action)
if action_id not in self._registered_actions:
await self._send_error_response(request_id, f"Action {action_id} not registered")
@@ -942,6 +1271,7 @@ class RTVIProcessor(FrameProcessor):
await self._push_transport_message(message)
async def _send_bot_ready(self):
"""Send the bot-ready message to the client."""
message = RTVIBotReady(
id=self._client_ready_id,
data=RTVIBotReadyData(version=RTVI_PROTOCOL_VERSION, config=self._config.config),
@@ -949,14 +1279,17 @@ class RTVIProcessor(FrameProcessor):
await self._push_transport_message(message)
async def _send_error_frame(self, frame: ErrorFrame):
"""Send an error frame as an RTVI error message."""
if self._errors_enabled:
message = RTVIError(data=RTVIErrorData(error=frame.error, fatal=frame.fatal))
await self._push_transport_message(message)
async def _send_error_response(self, id: str, error: str):
"""Send an error response message."""
if self._errors_enabled:
message = RTVIErrorResponse(id=id, data=RTVIErrorResponseData(error=error))
await self._push_transport_message(message)
def _action_id(self, service: str, action: str) -> str:
"""Generate an action ID from service and action names."""
return f"{service}:{action}"

41
src/pipecat/processors/gstreamer/pipeline_source.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""GStreamer pipeline source integration for Pipecat."""
import asyncio
from typing import Optional
@@ -36,7 +38,24 @@ except ModuleNotFoundError as e:
class GStreamerPipelineSource(FrameProcessor):
"""A frame processor that uses GStreamer pipelines as media sources.
This processor creates and manages GStreamer pipelines to generate audio and video
output frames. It handles pipeline lifecycle, decoding, format conversion, and
frame generation with configurable output parameters.
"""
class OutputParams(BaseModel):
"""Output configuration parameters for GStreamer pipeline.
Parameters:
video_width: Width of output video frames in pixels.
video_height: Height of output video frames in pixels.
audio_sample_rate: Sample rate for audio output. If None, uses frame sample rate.
audio_channels: Number of audio channels for output.
clock_sync: Whether to synchronize output with pipeline clock.
"""
video_width: int = 1280
video_height: int = 720
audio_sample_rate: Optional[int] = None
@@ -44,6 +63,13 @@ class GStreamerPipelineSource(FrameProcessor):
clock_sync: bool = True
def __init__(self, *, pipeline: str, out_params: Optional[OutputParams] = None, **kwargs):
"""Initialize the GStreamer pipeline source.
Args:
pipeline: GStreamer pipeline description string for the source.
out_params: Output configuration parameters. If None, uses defaults.
**kwargs: Additional arguments passed to parent FrameProcessor.
"""
super().__init__(**kwargs)
self._out_params = out_params or GStreamerPipelineSource.OutputParams()
@@ -67,6 +93,12 @@ class GStreamerPipelineSource(FrameProcessor):
bus.connect("message", self._on_gstreamer_message)
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and manage GStreamer pipeline lifecycle.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
# Specific system frames
@@ -92,13 +124,16 @@ class GStreamerPipelineSource(FrameProcessor):
await self.push_frame(frame, direction)
async def _start(self, frame: StartFrame):
"""Start the GStreamer pipeline."""
self._sample_rate = self._out_params.audio_sample_rate or frame.audio_out_sample_rate
self._player.set_state(Gst.State.PLAYING)
async def _stop(self, frame: EndFrame):
"""Stop the GStreamer pipeline."""
self._player.set_state(Gst.State.NULL)
async def _cancel(self, frame: CancelFrame):
"""Cancel the GStreamer pipeline."""
self._player.set_state(Gst.State.NULL)
#
@@ -106,6 +141,7 @@ class GStreamerPipelineSource(FrameProcessor):
#
def _on_gstreamer_message(self, bus: Gst.Bus, message: Gst.Message):
"""Handle GStreamer bus messages."""
t = message.type
if t == Gst.MessageType.ERROR:
err, debug = message.parse_error()
@@ -113,6 +149,7 @@ class GStreamerPipelineSource(FrameProcessor):
return True
def _decodebin_callback(self, decodebin: Gst.Element, pad: Gst.Pad):
"""Handle new pads from decodebin element."""
caps_string = pad.get_current_caps().to_string()
if caps_string.startswith("audio"):
self._decodebin_audio(pad)
@@ -120,6 +157,7 @@ class GStreamerPipelineSource(FrameProcessor):
self._decodebin_video(pad)
def _decodebin_audio(self, pad: Gst.Pad):
"""Set up audio processing pipeline from decoded audio pad."""
queue_audio = Gst.ElementFactory.make("queue", None)
audioconvert = Gst.ElementFactory.make("audioconvert", None)
audioresample = Gst.ElementFactory.make("audioresample", None)
@@ -153,6 +191,7 @@ class GStreamerPipelineSource(FrameProcessor):
pad.link(queue_pad)
def _decodebin_video(self, pad: Gst.Pad):
"""Set up video processing pipeline from decoded video pad."""
queue_video = Gst.ElementFactory.make("queue", None)
videoconvert = Gst.ElementFactory.make("videoconvert", None)
videoscale = Gst.ElementFactory.make("videoscale", None)
@@ -187,6 +226,7 @@ class GStreamerPipelineSource(FrameProcessor):
pad.link(queue_pad)
def _appsink_audio_new_sample(self, appsink: GstApp.AppSink):
"""Handle new audio samples from GStreamer appsink."""
buffer = appsink.pull_sample().get_buffer()
(_, info) = buffer.map(Gst.MapFlags.READ)
frame = OutputAudioRawFrame(
@@ -199,6 +239,7 @@ class GStreamerPipelineSource(FrameProcessor):
return Gst.FlowReturn.OK
def _appsink_video_new_sample(self, appsink: GstApp.AppSink):
"""Handle new video samples from GStreamer appsink."""
buffer = appsink.pull_sample().get_buffer()
(_, info) = buffer.map(Gst.MapFlags.READ)
frame = OutputImageRawFrame(

29
src/pipecat/processors/idle_frame_processor.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Idle frame processor for timeout-based callback execution."""
import asyncio
from typing import Awaitable, Callable, List, Optional
@@ -12,9 +14,11 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class IdleFrameProcessor(FrameProcessor):
"""This class waits to receive any frame or list of desired frames within a
given timeout. If the timeout is reached before receiving any of those
frames the provided callback will be called.
"""Monitors frame activity and triggers callbacks on timeout.
This processor waits to receive any frame or specific frame types within a
given timeout period. If the timeout is reached before receiving the expected
frames, the provided callback will be executed.
"""
def __init__(
@@ -25,6 +29,16 @@ class IdleFrameProcessor(FrameProcessor):
types: Optional[List[type]] = None,
**kwargs,
):
"""Initialize the idle frame processor.
Args:
callback: Async callback function to execute on timeout. Receives
this processor instance as an argument.
timeout: Timeout duration in seconds before triggering the callback.
types: Optional list of frame types to monitor. If None, monitors
all frames.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._callback = callback
@@ -33,6 +47,12 @@ class IdleFrameProcessor(FrameProcessor):
self._idle_task = None
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and manage idle timeout monitoring.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
@@ -50,15 +70,18 @@ class IdleFrameProcessor(FrameProcessor):
self._idle_event.set()
async def cleanup(self):
"""Clean up resources and cancel pending tasks."""
if self._idle_task:
await self.cancel_task(self._idle_task)
def _create_idle_task(self):
"""Create and start the idle monitoring task."""
if not self._idle_task:
self._idle_event = asyncio.Event()
self._idle_task = self.create_task(self._idle_task_handler())
async def _idle_task_handler(self):
"""Handle idle timeout monitoring and callback execution."""
while True:
try:
await asyncio.wait_for(self._idle_event.wait(), timeout=self._timeout)

23
src/pipecat/processors/logger.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Frame logging utilities for debugging and monitoring frame flow in Pipecat pipelines."""
from typing import Optional, Tuple, Type
from loguru import logger
@@ -21,6 +23,13 @@ logger = logger.opt(ansi=True)
class FrameLogger(FrameProcessor):
"""A frame processor that logs frame information for debugging purposes.
This processor intercepts frames passing through the pipeline and logs
their details with configurable formatting and filtering. Useful for
debugging frame flow and understanding pipeline behavior.
"""
def __init__(
self,
prefix="Frame",
@@ -32,12 +41,26 @@ class FrameLogger(FrameProcessor):
TransportMessageFrame,
),
):
"""Initialize the frame logger.
Args:
prefix: Text prefix to add to log messages. Defaults to "Frame".
color: ANSI color code for log message formatting. If None, no coloring is applied.
ignored_frame_types: Tuple of frame types to exclude from logging.
Defaults to common high-frequency frames like audio and speaking frames.
"""
super().__init__()
self._prefix = prefix
self._color = color
self._ignored_frame_types = ignored_frame_types
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process and log frame information.
Args:
frame: The frame to process and potentially log.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if self._ignored_frame_types and not isinstance(frame, self._ignored_frame_types):

87
src/pipecat/processors/metrics/frame_processor_metrics.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Frame processor metrics collection and reporting."""
import time
from typing import Optional
@@ -18,12 +20,25 @@ from pipecat.metrics.metrics import (
TTFBMetricsData,
TTSUsageMetricsData,
)
from pipecat.utils.asyncio import TaskManager
from pipecat.utils.asyncio.task_manager import BaseTaskManager
from pipecat.utils.base_object import BaseObject
class FrameProcessorMetrics(BaseObject):
"""Metrics collection and reporting for frame processors.
Provides comprehensive metrics tracking for frame processing operations,
including timing measurements, resource usage, and performance analytics.
Supports TTFB tracking, processing duration metrics, and usage statistics
for LLM and TTS operations.
"""
def __init__(self):
"""Initialize the frame processor metrics collector.
Sets up internal state for tracking various metrics including TTFB,
processing times, and usage statistics.
"""
super().__init__()
self._task_manager = None
self._start_ttfb_time = 0
@@ -31,14 +46,25 @@ class FrameProcessorMetrics(BaseObject):
self._last_ttfb_time = 0
self._should_report_ttfb = True
async def setup(self, task_manager: TaskManager):
async def setup(self, task_manager: BaseTaskManager):
"""Set up the metrics collector with a task manager.
Args:
task_manager: The task manager for handling async operations.
"""
self._task_manager = task_manager
async def cleanup(self):
"""Clean up metrics collection resources."""
await super().cleanup()
@property
def task_manager(self) -> TaskManager:
def task_manager(self) -> BaseTaskManager:
"""Get the associated task manager.
Returns:
The task manager instance for async operations.
"""
return self._task_manager
@property
@@ -46,7 +72,7 @@ class FrameProcessorMetrics(BaseObject):
"""Get the current TTFB value in seconds.
Returns:
Optional[float]: The TTFB value in seconds, or None if not measured
The TTFB value in seconds, or None if not measured.
"""
if self._last_ttfb_time > 0:
return self._last_ttfb_time
@@ -58,24 +84,46 @@ class FrameProcessorMetrics(BaseObject):
return None
def _processor_name(self):
"""Get the processor name from core metrics data."""
return self._core_metrics_data.processor
def _model_name(self):
"""Get the model name from core metrics data."""
return self._core_metrics_data.model
def set_core_metrics_data(self, data: MetricsData):
"""Set the core metrics data for this collector.
Args:
data: The core metrics data containing processor and model information.
"""
self._core_metrics_data = data
def set_processor_name(self, name: str):
"""Set the processor name for metrics reporting.
Args:
name: The name of the processor to use in metrics.
"""
self._core_metrics_data = MetricsData(processor=name)
async def start_ttfb_metrics(self, report_only_initial_ttfb):
"""Start measuring time-to-first-byte (TTFB).
Args:
report_only_initial_ttfb: Whether to report only the first TTFB measurement.
"""
if self._should_report_ttfb:
self._start_ttfb_time = time.time()
self._last_ttfb_time = 0
self._should_report_ttfb = not report_only_initial_ttfb
async def stop_ttfb_metrics(self):
"""Stop TTFB measurement and generate metrics frame.
Returns:
MetricsFrame containing TTFB data, or None if not measuring.
"""
if self._start_ttfb_time == 0:
return None
@@ -88,9 +136,15 @@ class FrameProcessorMetrics(BaseObject):
return MetricsFrame(data=[ttfb])
async def start_processing_metrics(self):
"""Start measuring processing time."""
self._start_processing_time = time.time()
async def stop_processing_metrics(self):
"""Stop processing time measurement and generate metrics frame.
Returns:
MetricsFrame containing processing duration data, or None if not measuring.
"""
if self._start_processing_time == 0:
return None
@@ -103,15 +157,34 @@ class FrameProcessorMetrics(BaseObject):
return MetricsFrame(data=[processing])
async def start_llm_usage_metrics(self, tokens: LLMTokenUsage):
logger.debug(
f"{self._processor_name()} prompt tokens: {tokens.prompt_tokens}, completion tokens: {tokens.completion_tokens}"
)
"""Record LLM token usage metrics.
Args:
tokens: Token usage information including prompt and completion tokens.
Returns:
MetricsFrame containing LLM usage data.
"""
logstr = f"{self._processor_name()} prompt tokens: {tokens.prompt_tokens}, completion tokens: {tokens.completion_tokens}"
if tokens.cache_read_input_tokens:
logstr += f", cache read input tokens: {tokens.cache_read_input_tokens}"
if tokens.reasoning_tokens:
logstr += f", reasoning tokens: {tokens.reasoning_tokens}"
logger.debug(logstr)
value = LLMUsageMetricsData(
processor=self._processor_name(), model=self._model_name(), value=tokens
)
return MetricsFrame(data=[value])
async def start_tts_usage_metrics(self, text: str):
"""Record TTS character usage metrics.
Args:
text: The text being processed by TTS.
Returns:
MetricsFrame containing TTS usage data.
"""
characters = TTSUsageMetricsData(
processor=self._processor_name(), model=self._model_name(), value=len(text)
)

50
src/pipecat/processors/metrics/sentry.py
View File

@@ -4,11 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
import asyncio
"""Sentry integration for frame processor metrics."""
from loguru import logger
from pipecat.utils.asyncio import TaskManager
from pipecat.utils.asyncio.task_manager import BaseTaskManager
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
try:
import sentry_sdk
@@ -21,25 +22,45 @@ from pipecat.processors.metrics.frame_processor_metrics import FrameProcessorMet
class SentryMetrics(FrameProcessorMetrics):
"""Frame processor metrics integration with Sentry monitoring.
Extends FrameProcessorMetrics to send time-to-first-byte (TTFB) and
processing metrics as Sentry transactions for performance monitoring
and debugging.
"""
def __init__(self):
"""Initialize the Sentry metrics collector.
Sets up internal state for tracking transactions and verifies
Sentry SDK initialization status.
"""
super().__init__()
self._ttfb_metrics_tx = None
self._processing_metrics_tx = None
self._sentry_available = sentry_sdk.is_initialized()
if not self._sentry_available:
logger.warning("Sentry SDK not initialized. Sentry features will be disabled.")
self._sentry_queue = asyncio.Queue()
self._sentry_task = None
async def setup(self, task_manager: TaskManager):
async def setup(self, task_manager: BaseTaskManager):
"""Setup the Sentry metrics system.
Args:
task_manager: The task manager to use for background operations.
"""
await super().setup(task_manager)
if self._sentry_available:
self._sentry_queue = asyncio.Queue()
self._sentry_queue = WatchdogQueue(task_manager)
self._sentry_task = self.task_manager.create_task(
self._sentry_task_handler(), name=f"{self}::_sentry_task_handler"
)
async def cleanup(self):
"""Clean up Sentry resources and flush pending transactions.
Ensures all pending transactions are sent to Sentry before shutdown.
"""
await super().cleanup()
if self._sentry_task:
await self._sentry_queue.put(None)
@@ -49,6 +70,11 @@ class SentryMetrics(FrameProcessorMetrics):
sentry_sdk.flush(timeout=5.0)
async def start_ttfb_metrics(self, report_only_initial_ttfb):
"""Start tracking time-to-first-byte metrics.
Args:
report_only_initial_ttfb: Whether to report only the initial TTFB measurement.
"""
await super().start_ttfb_metrics(report_only_initial_ttfb)
if self._should_report_ttfb and self._sentry_available:
@@ -61,6 +87,10 @@ class SentryMetrics(FrameProcessorMetrics):
)
async def stop_ttfb_metrics(self):
"""Stop tracking time-to-first-byte metrics.
Queues the TTFB transaction for completion and transmission to Sentry.
"""
await super().stop_ttfb_metrics()
if self._sentry_available and self._ttfb_metrics_tx:
@@ -68,6 +98,10 @@ class SentryMetrics(FrameProcessorMetrics):
self._ttfb_metrics_tx = None
async def start_processing_metrics(self):
"""Start tracking frame processing metrics.
Creates a new Sentry transaction to track processing performance.
"""
await super().start_processing_metrics()
if self._sentry_available:
@@ -80,6 +114,10 @@ class SentryMetrics(FrameProcessorMetrics):
)
async def stop_processing_metrics(self):
"""Stop tracking frame processing metrics.
Queues the processing transaction for completion and transmission to Sentry.
"""
await super().stop_processing_metrics()
if self._sentry_available and self._processing_metrics_tx:
@@ -87,9 +125,11 @@ class SentryMetrics(FrameProcessorMetrics):
self._processing_metrics_tx = None
async def _sentry_task_handler(self):
"""Background task handler for completing Sentry transactions."""
running = True
while running:
tx = await self._sentry_queue.get()
if tx:
await self.task_manager.get_event_loop().run_in_executor(None, tx.finish)
running = tx is not None
self._sentry_queue.task_done()

48
src/pipecat/processors/producer_processor.py
View File

@@ -4,23 +4,35 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Producer processor for frame filtering and distribution."""
import asyncio
from typing import Awaitable, Callable, List
from pipecat.frames.frames import Frame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.utils.asyncio.watchdog_queue import WatchdogQueue
async def identity_transformer(frame: Frame):
"""Default transformer that returns the frame unchanged.
Args:
frame: The frame to transform.
Returns:
The same frame without modifications.
"""
return frame
class ProducerProcessor(FrameProcessor):
"""This class optionally passes-through received frames and decides if those
frames should be sent to consumers based on a user-defined filter. The
frames can be transformed into a different type of frame before being
sending them to the consumers. More than one consumer can be added.
"""A processor that filters frames and distributes them to multiple consumers.
This processor receives frames, applies a filter to determine which frames
should be sent to consumers (ConsumerProcessor), optionally transforms those
frames, and distributes them to registered consumer queues. It can also pass
frames through to the next processor in the pipeline.
"""
def __init__(
@@ -30,6 +42,16 @@ class ProducerProcessor(FrameProcessor):
transformer: Callable[[Frame], Awaitable[Frame]] = identity_transformer,
passthrough: bool = True,
):
"""Initialize the producer processor.
Args:
filter: Async function that determines if a frame should be produced.
Must return True for frames to be sent to consumers.
transformer: Async function to transform frames before sending to consumers.
Defaults to identity_transformer which returns frames unchanged.
passthrough: Whether to pass frames through to the next processor.
If True, all frames continue downstream regardless of filter result.
"""
super().__init__()
self._filter = filter
self._transformer = transformer
@@ -37,26 +59,25 @@ class ProducerProcessor(FrameProcessor):
self._consumers: List[asyncio.Queue] = []
def add_consumer(self):
"""
Adds a new consumer and returns its associated queue.
"""Add a new consumer and return its associated queue.
Returns:
asyncio.Queue: The queue for the newly added consumer.
"""
queue = asyncio.Queue()
queue = WatchdogQueue(self.task_manager)
self._consumers.append(queue)
return queue
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""
Processes an incoming frame and determines whether to produce it as a ProducerItem.
"""Process an incoming frame and determine whether to produce it.
If the frame meets the produce criteria, it will be added to the consumer queues.
If passthrough is enabled, the frame will also be sent to consumers.
If the frame meets the filter criteria, it will be transformed and added
to all consumer queues. If passthrough is enabled, the original frame
will also be sent downstream.
Args:
frame (Frame): The frame to process.
direction (FrameDirection): The direction of the frame.
frame: The frame to process.
direction: The direction of the frame flow.
"""
await super().process_frame(frame, direction)
@@ -68,6 +89,7 @@ class ProducerProcessor(FrameProcessor):
await self.push_frame(frame, direction)
async def _produce(self, frame: Frame):
"""Produce a frame to all consumers."""
for consumer in self._consumers:
new_frame = await self._transformer(frame)
await consumer.put(new_frame)

26
src/pipecat/processors/text_transformer.py
View File

@@ -4,14 +4,20 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
from typing import Coroutine
"""Stateless text transformation processor for Pipecat."""
from typing import Callable, Coroutine, Union
from pipecat.frames.frames import Frame, TextFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class StatelessTextTransformer(FrameProcessor):
"""This processor calls the given function on any text in a text frame.
"""Processor that applies transformation functions to text frames.
This processor intercepts TextFrame objects and applies a user-provided
transformation function to the text content. The function can be either
synchronous or asynchronous (coroutine).
>>> async def print_frames(aggregator, frame):
... async for frame in aggregator.process_frame(frame):
@@ -22,11 +28,25 @@ class StatelessTextTransformer(FrameProcessor):
HELLO
"""
def __init__(self, transform_fn):
def __init__(
self, transform_fn: Union[Callable[[str], str], Callable[[str], Coroutine[None, None, str]]]
):
"""Initialize the text transformer.
Args:
transform_fn: Function to apply to text content. Can be synchronous
(str -> str) or asynchronous (str -> Coroutine[None, None, str]).
"""
super().__init__()
self._transform_fn = transform_fn
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames, applying transformation to text frames.
Args:
frame: The frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, TextFrame):

46
src/pipecat/processors/transcript_processor.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Transcript processing utilities for conversation recording and analysis.
This module provides processors that convert speech and text frames into structured
transcript messages with timestamps, enabling conversation history tracking and analysis.
"""
from typing import List, Optional
from loguru import logger
@@ -30,7 +36,11 @@ class BaseTranscriptProcessor(FrameProcessor):
"""
def __init__(self, **kwargs):
"""Initialize processor with empty message store."""
"""Initialize processor with empty message store.
Args:
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._processed_messages: List[TranscriptionMessage] = []
self._register_event_handler("on_transcript_update")
@@ -39,7 +49,7 @@ class BaseTranscriptProcessor(FrameProcessor):
"""Emit transcript updates for new messages.
Args:
messages: New messages to emit in update
messages: New messages to emit in update.
"""
if messages:
self._processed_messages.extend(messages)
@@ -55,8 +65,8 @@ class UserTranscriptProcessor(BaseTranscriptProcessor):
"""Process TranscriptionFrames into user conversation messages.
Args:
frame: Input frame to process
direction: Frame processing direction
frame: Input frame to process.
direction: Frame processing direction.
"""
await super().process_frame(frame, direction)
@@ -77,14 +87,14 @@ class AssistantTranscriptProcessor(BaseTranscriptProcessor):
- The bot stops speaking (BotStoppedSpeakingFrame)
- The bot is interrupted (StartInterruptionFrame)
- The pipeline ends (EndFrame)
Attributes:
_current_text_parts: List of text fragments being aggregated for current utterance
_aggregation_start_time: Timestamp when the current utterance began
"""
def __init__(self, **kwargs):
"""Initialize processor with aggregation state."""
"""Initialize processor with aggregation state.
Args:
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
self._current_text_parts: List[str] = []
self._aggregation_start_time: Optional[str] = None
@@ -176,8 +186,8 @@ class AssistantTranscriptProcessor(BaseTranscriptProcessor):
- CancelFrame: Completes current utterance due to cancellation
Args:
frame: Input frame to process
direction: Frame processing direction
frame: Input frame to process.
direction: Frame processing direction.
"""
await super().process_frame(frame, direction)
@@ -245,7 +255,10 @@ class TranscriptProcessor:
"""Get the user transcript processor.
Args:
**kwargs: Arguments specific to UserTranscriptProcessor
**kwargs: Arguments specific to UserTranscriptProcessor.
Returns:
The user transcript processor instance.
"""
if self._user_processor is None:
self._user_processor = UserTranscriptProcessor(**kwargs)
@@ -262,7 +275,10 @@ class TranscriptProcessor:
"""Get the assistant transcript processor.
Args:
**kwargs: Arguments specific to AssistantTranscriptProcessor
**kwargs: Arguments specific to AssistantTranscriptProcessor.
Returns:
The assistant transcript processor instance.
"""
if self._assistant_processor is None:
self._assistant_processor = AssistantTranscriptProcessor(**kwargs)
@@ -279,10 +295,10 @@ class TranscriptProcessor:
"""Register event handler for both processors.
Args:
event_name: Name of event to handle
event_name: Name of event to handle.
Returns:
Decorator function that registers handler with both processors
Decorator function that registers handler with both processors.
"""
def decorator(handler):

39
src/pipecat/processors/user_idle_processor.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""User idle detection and timeout handling for Pipecat."""
import asyncio
import inspect
from typing import Awaitable, Callable, Union
@@ -22,19 +24,12 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class UserIdleProcessor(FrameProcessor):
"""Monitors user inactivity and triggers callbacks after timeout periods.
Starts monitoring only after the first conversation activity (UserStartedSpeaking
or BotSpeaking).
Args:
callback: Function to call when user is idle. Can be either:
- Basic callback(processor) -> None
- Retry callback(processor, retry_count) -> bool
Return True to continue monitoring for idle events,
Return False to stop the idle monitoring task
timeout: Seconds to wait before considering user idle
**kwargs: Additional arguments passed to FrameProcessor
This processor tracks user activity and triggers configurable callbacks when
users become idle. It starts monitoring only after the first conversation
activity and supports both basic and retry-based callback patterns.
Example:
```
# Retry callback:
async def handle_idle(processor: "UserIdleProcessor", retry_count: int) -> bool:
if retry_count < 3:
@@ -50,6 +45,7 @@ class UserIdleProcessor(FrameProcessor):
callback=handle_idle,
timeout=5.0
)
```
"""
def __init__(
@@ -62,6 +58,17 @@ class UserIdleProcessor(FrameProcessor):
timeout: float,
**kwargs,
):
"""Initialize the user idle processor.
Args:
callback: Function to call when user is idle. Can be either:
- Basic callback(processor) -> None
- Retry callback(processor, retry_count) -> bool
Return True to continue monitoring for idle events,
Return False to stop the idle monitoring task
timeout: Seconds to wait before considering user idle.
**kwargs: Additional arguments passed to FrameProcessor.
"""
super().__init__(**kwargs)
self._callback = self._wrap_callback(callback)
self._timeout = timeout
@@ -107,7 +114,11 @@ class UserIdleProcessor(FrameProcessor):
@property
def retry_count(self) -> int:
"""Returns the current retry count."""
"""Get the current retry count.
Returns:
The number of times the idle callback has been triggered.
"""
return self._retry_count
async def _stop(self) -> None:
@@ -120,8 +131,8 @@ class UserIdleProcessor(FrameProcessor):
"""Processes incoming frames and manages idle monitoring state.
Args:
frame: The frame to process
direction: Direction of the frame flow
frame: The frame to process.
direction: Direction of the frame flow.
"""
await super().process_frame(frame, direction)

70
src/pipecat/services/ai_service.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Base AI service implementation.
Provides the foundation for all AI services in the Pipecat framework, including
model management, settings handling, and frame processing lifecycle methods.
"""
from typing import Any, AsyncGenerator, Dict, Mapping
from loguru import logger
@@ -20,7 +26,20 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
class AIService(FrameProcessor):
"""Base class for all AI services.
Provides common functionality for AI services including model management,
settings handling, session properties, and frame processing lifecycle.
Subclasses should implement specific AI functionality while leveraging
this base infrastructure.
"""
def __init__(self, **kwargs):
"""Initialize the AI service.
Args:
**kwargs: Additional arguments passed to the parent FrameProcessor.
"""
super().__init__(**kwargs)
self._model_name: str = ""
self._settings: Dict[str, Any] = {}
@@ -28,19 +47,53 @@ class AIService(FrameProcessor):
@property
def model_name(self) -> str:
"""Get the current model name.
Returns:
The name of the AI model being used.
"""
return self._model_name
def set_model_name(self, model: str):
"""Set the AI model name and update metrics.
Args:
model: The name of the AI model to use.
"""
self._model_name = model
self.set_core_metrics_data(MetricsData(processor=self.name, model=self._model_name))
async def start(self, frame: StartFrame):
"""Start the AI service.
Called when the service should begin processing. Subclasses should
override this method to perform service-specific initialization.
Args:
frame: The start frame containing initialization parameters.
"""
pass
async def stop(self, frame: EndFrame):
"""Stop the AI service.
Called when the service should stop processing. Subclasses should
override this method to perform cleanup operations.
Args:
frame: The end frame.
"""
pass
async def cancel(self, frame: CancelFrame):
"""Cancel the AI service.
Called when the service should cancel all operations. Subclasses should
override this method to handle cancellation logic.
Args:
frame: The cancel frame.
"""
pass
async def _update_settings(self, settings: Mapping[str, Any]):
@@ -87,6 +140,15 @@ class AIService(FrameProcessor):
logger.warning(f"Unknown setting for {self.name} service: {key}")
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames and handle service lifecycle.
Automatically handles StartFrame, EndFrame, and CancelFrame by calling
the appropriate lifecycle methods.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartFrame):
@@ -97,6 +159,14 @@ class AIService(FrameProcessor):
await self.stop(frame)
async def process_generator(self, generator: AsyncGenerator[Frame | None, None]):
"""Process frames from an async generator.
Takes an async generator that yields frames and processes each one,
handling error frames specially by pushing them as errors.
Args:
generator: An async generator that yields Frame objects or None.
"""
async for f in generator:
if f:
if isinstance(f, ErrorFrame):

11
src/pipecat/services/ai_services.py
View File

@@ -4,6 +4,17 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Deprecated AI services module.
This module is deprecated. Import services directly from their respective modules:
- pipecat.services.ai_service
- pipecat.services.image_service
- pipecat.services.llm_service
- pipecat.services.stt_service
- pipecat.services.tts_service
- pipecat.services.vision_service
"""
import sys
from pipecat.services import DeprecatedModuleProxy

233
src/pipecat/services/anthropic/llm.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Anthropic AI service integration for Pipecat.
This module provides LLM services and context management for Anthropic's Claude models,
including support for function calling, vision, and prompt caching features.
"""
import asyncio
import base64
import copy
@@ -46,6 +52,7 @@ from pipecat.processors.aggregators.openai_llm_context import (
)
from pipecat.processors.frame_processor import FrameDirection
from pipecat.services.llm_service import FunctionCallFromLLM, LLMService
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.tracing.service_decorators import traced_llm
try:
@@ -58,27 +65,59 @@ except ModuleNotFoundError as e:
@dataclass
class AnthropicContextAggregatorPair:
"""Pair of context aggregators for Anthropic conversations.
Encapsulates both user and assistant context aggregators
to manage conversation flow and message formatting.
Parameters:
_user: The user context aggregator.
_assistant: The assistant context aggregator.
"""
_user: "AnthropicUserContextAggregator"
_assistant: "AnthropicAssistantContextAggregator"
def user(self) -> "AnthropicUserContextAggregator":
"""Get the user context aggregator.
Returns:
The user context aggregator instance.
"""
return self._user
def assistant(self) -> "AnthropicAssistantContextAggregator":
"""Get the assistant context aggregator.
Returns:
The assistant context aggregator instance.
"""
return self._assistant
class AnthropicLLMService(LLMService):
"""This class implements inference with Anthropic's AI models.
"""LLM service for Anthropic's Claude models.
Can provide a custom client via the `client` kwarg, allowing you to
use `AsyncAnthropicBedrock` and `AsyncAnthropicVertex` clients
Provides inference capabilities with Claude models including support for
function calling, vision processing, streaming responses, and prompt caching.
Can use custom clients like AsyncAnthropicBedrock and AsyncAnthropicVertex.
"""
# Overriding the default adapter to use the Anthropic one.
adapter_class = AnthropicLLMAdapter
class InputParams(BaseModel):
"""Input parameters for Anthropic model inference.
Parameters:
enable_prompt_caching_beta: Whether to enable beta prompt caching feature.
max_tokens: Maximum tokens to generate. Must be at least 1.
temperature: Sampling temperature between 0.0 and 1.0.
top_k: Top-k sampling parameter.
top_p: Top-p sampling parameter between 0.0 and 1.0.
extra: Additional parameters to pass to the API.
"""
enable_prompt_caching_beta: Optional[bool] = False
max_tokens: Optional[int] = Field(default_factory=lambda: 4096, ge=1)
temperature: Optional[float] = Field(default_factory=lambda: NOT_GIVEN, ge=0.0, le=1.0)
@@ -95,6 +134,15 @@ class AnthropicLLMService(LLMService):
client=None,
**kwargs,
):
"""Initialize the Anthropic LLM service.
Args:
api_key: Anthropic API key for authentication.
model: Model name to use. Defaults to "claude-sonnet-4-20250514".
params: Optional model parameters for inference.
client: Optional custom Anthropic client instance.
**kwargs: Additional arguments passed to parent LLMService.
"""
super().__init__(**kwargs)
params = params or AnthropicLLMService.InputParams()
self._client = client or AsyncAnthropic(
@@ -111,10 +159,20 @@ class AnthropicLLMService(LLMService):
}
def can_generate_metrics(self) -> bool:
"""Check if this service can generate usage metrics.
Returns:
True, as Anthropic provides detailed token usage metrics.
"""
return True
@property
def enable_prompt_caching_beta(self) -> bool:
"""Check if prompt caching beta feature is enabled.
Returns:
True if prompt caching is enabled.
"""
return self._enable_prompt_caching_beta
def create_context_aggregator(
@@ -124,22 +182,19 @@ class AnthropicLLMService(LLMService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> AnthropicContextAggregatorPair:
"""Create an instance of AnthropicContextAggregatorPair from an
OpenAILLMContext. Constructor keyword arguments for both the user and
assistant aggregators can be provided.
"""Create Anthropic-specific context aggregators.
Creates a pair of context aggregators optimized for Anthropic's message format,
including support for function calls, tool usage, and image handling.
Args:
context (OpenAILLMContext): The LLM context.
user_params (LLMUserAggregatorParams, optional): User aggregator
parameters.
assistant_params (LLMAssistantAggregatorParams, optional): User
aggregator parameters.
context: The LLM context.
user_params: User aggregator parameters.
assistant_params: Assistant aggregator parameters.
Returns:
AnthropicContextAggregatorPair: A pair of context aggregators, one
for the user and one for the assistant, encapsulated in an
AnthropicContextAggregatorPair.
A pair of context aggregators, one for the user and one for the assistant,
encapsulated in an AnthropicContextAggregatorPair.
"""
context.set_llm_adapter(self.get_llm_adapter())
@@ -203,7 +258,7 @@ class AnthropicLLMService(LLMService):
json_accumulator = ""
function_calls = []
async for event in response:
async for event in WatchdogAsyncIterator(response, manager=self.task_manager):
# Aggregate streaming content, create frames, trigger events
if event.type == "content_block_delta":
@@ -307,6 +362,15 @@ class AnthropicLLMService(LLMService):
)
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and route them appropriately.
Handles various frame types including context frames, message frames,
vision frames, and settings updates.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
context = None
@@ -358,6 +422,13 @@ class AnthropicLLMService(LLMService):
class AnthropicLLMContext(OpenAILLMContext):
"""LLM context specialized for Anthropic's message format and features.
Extends OpenAILLMContext to handle Anthropic-specific features like
system messages, prompt caching, and message format conversions.
Manages conversation state and message history formatting.
"""
def __init__(
self,
messages: Optional[List[dict]] = None,
@@ -366,6 +437,14 @@ class AnthropicLLMContext(OpenAILLMContext):
*,
system: Union[str, NotGiven] = NOT_GIVEN,
):
"""Initialize the Anthropic LLM context.
Args:
messages: Initial list of conversation messages.
tools: Available function calling tools.
tool_choice: Tool selection preference.
system: System message content.
"""
super().__init__(messages=messages, tools=tools, tool_choice=tool_choice)
# For beta prompt caching. This is a counter that tracks the number of turns
@@ -378,6 +457,16 @@ class AnthropicLLMContext(OpenAILLMContext):
@staticmethod
def upgrade_to_anthropic(obj: OpenAILLMContext) -> "AnthropicLLMContext":
"""Upgrade an OpenAI context to Anthropic format.
Converts message format and restructures content for Anthropic compatibility.
Args:
obj: The OpenAI context to upgrade.
Returns:
The upgraded Anthropic context.
"""
logger.debug(f"Upgrading to Anthropic: {obj}")
if isinstance(obj, OpenAILLMContext) and not isinstance(obj, AnthropicLLMContext):
obj.__class__ = AnthropicLLMContext
@@ -386,6 +475,14 @@ class AnthropicLLMContext(OpenAILLMContext):
@classmethod
def from_openai_context(cls, openai_context: OpenAILLMContext):
"""Create Anthropic context from OpenAI context.
Args:
openai_context: The OpenAI context to convert.
Returns:
New Anthropic context with converted messages.
"""
self = cls(
messages=openai_context.messages,
tools=openai_context.tools,
@@ -397,12 +494,28 @@ class AnthropicLLMContext(OpenAILLMContext):
@classmethod
def from_messages(cls, messages: List[dict]) -> "AnthropicLLMContext":
"""Create context from a list of messages.
Args:
messages: List of conversation messages.
Returns:
New Anthropic context with the provided messages.
"""
self = cls(messages=messages)
self._restructure_from_openai_messages()
return self
@classmethod
def from_image_frame(cls, frame: VisionImageRawFrame) -> "AnthropicLLMContext":
"""Create context from a vision image frame.
Args:
frame: The vision image frame to process.
Returns:
New Anthropic context with the image message.
"""
context = cls()
context.add_image_frame_message(
format=frame.format, size=frame.size, image=frame.image, text=frame.text
@@ -410,11 +523,15 @@ class AnthropicLLMContext(OpenAILLMContext):
return context
def set_messages(self, messages: List):
"""Set the messages list and reset cache tracking.
Args:
messages: New list of messages to set.
"""
self.turns_above_cache_threshold = 0
self._messages[:] = messages
self._restructure_from_openai_messages()
# convert a message in Anthropic format into one or more messages in OpenAI format
def to_standard_messages(self, obj):
"""Convert Anthropic message format to standard structured format.
@@ -555,6 +672,17 @@ class AnthropicLLMContext(OpenAILLMContext):
def add_image_frame_message(
self, *, format: str, size: tuple[int, int], image: bytes, text: str = None
):
"""Add an image message to the context.
Converts the image to base64 JPEG format and adds it as a user message
with optional accompanying text.
Args:
format: The image format (e.g., 'RGB', 'RGBA').
size: Image dimensions as (width, height).
image: Raw image bytes.
text: Optional text to accompany the image.
"""
buffer = io.BytesIO()
Image.frombytes(format, size, image).save(buffer, format="JPEG")
encoded_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
@@ -575,6 +703,14 @@ class AnthropicLLMContext(OpenAILLMContext):
self.add_message({"role": "user", "content": content})
def add_message(self, message):
"""Add a message to the context, merging with previous message if same role.
Anthropic requires alternating roles, so consecutive messages from the same
role are merged together.
Args:
message: The message to add to the context.
"""
try:
if self.messages:
# Anthropic requires that roles alternate. If this message's role is the same as the
@@ -600,6 +736,14 @@ class AnthropicLLMContext(OpenAILLMContext):
logger.error(f"Error adding message: {e}")
def get_messages_with_cache_control_markers(self) -> List[dict]:
"""Get messages with prompt caching markers applied.
Adds cache control markers to appropriate messages based on the
number of turns above the cache threshold.
Returns:
List of messages with cache control markers added.
"""
try:
messages = copy.deepcopy(self.messages)
if self.turns_above_cache_threshold >= 1 and messages[-1]["role"] == "user":
@@ -667,12 +811,26 @@ class AnthropicLLMContext(OpenAILLMContext):
message["content"] = [{"type": "text", "text": "(empty)"}]
def get_messages_for_persistent_storage(self):
"""Get messages formatted for persistent storage.
Includes system message at the beginning if present.
Returns:
List of messages suitable for storage.
"""
messages = super().get_messages_for_persistent_storage()
if self.system:
messages.insert(0, {"role": "system", "content": self.system})
return messages
def get_messages_for_logging(self) -> str:
"""Get messages formatted for logging with sensitive data redacted.
Replaces image data with placeholder text for cleaner logs.
Returns:
JSON string representation of messages for logging.
"""
msgs = []
for message in self.messages:
msg = copy.deepcopy(message)
@@ -686,6 +844,12 @@ class AnthropicLLMContext(OpenAILLMContext):
class AnthropicUserContextAggregator(LLMUserContextAggregator):
"""Anthropic-specific user context aggregator.
Handles aggregation of user messages for Anthropic LLM services.
Inherits all functionality from the base LLMUserContextAggregator.
"""
pass
@@ -700,7 +864,20 @@ class AnthropicUserContextAggregator(LLMUserContextAggregator):
class AnthropicAssistantContextAggregator(LLMAssistantContextAggregator):
"""Context aggregator for assistant messages in Anthropic conversations.
Handles function call lifecycle management including in-progress tracking,
result handling, and cancellation for Anthropic's tool use format.
"""
async def handle_function_call_in_progress(self, frame: FunctionCallInProgressFrame):
"""Handle a function call that is starting.
Creates tool use message and placeholder tool result for tracking.
Args:
frame: Frame containing function call details.
"""
assistant_message = {"role": "assistant", "content": []}
assistant_message["content"].append(
{
@@ -725,6 +902,13 @@ class AnthropicAssistantContextAggregator(LLMAssistantContextAggregator):
)
async def handle_function_call_result(self, frame: FunctionCallResultFrame):
"""Handle the result of a completed function call.
Updates the tool result with actual return value or completion status.
Args:
frame: Frame containing function call result.
"""
if frame.result:
result = json.dumps(frame.result)
await self._update_function_call_result(frame.function_name, frame.tool_call_id, result)
@@ -734,6 +918,13 @@ class AnthropicAssistantContextAggregator(LLMAssistantContextAggregator):
)
async def handle_function_call_cancel(self, frame: FunctionCallCancelFrame):
"""Handle cancellation of a function call.
Updates the tool result to indicate cancellation.
Args:
frame: Frame containing function call cancellation details.
"""
await self._update_function_call_result(
frame.function_name, frame.tool_call_id, "CANCELLED"
)
@@ -752,6 +943,14 @@ class AnthropicAssistantContextAggregator(LLMAssistantContextAggregator):
content["content"] = result
async def handle_user_image_frame(self, frame: UserImageRawFrame):
"""Handle a user image frame with function call context.
Marks the associated function call as completed and adds the image
to the conversation context.
Args:
frame: User image frame with request context.
"""
await self._update_function_call_result(
frame.request.function_name, frame.request.tool_call_id, "COMPLETED"
)

68
src/pipecat/services/assemblyai/models.py
View File

@@ -1,10 +1,30 @@
#
# Copyright (c) 2024–2025, Daily
#
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AssemblyAI WebSocket API message models and connection parameters.
This module defines Pydantic models for handling AssemblyAI's real-time
transcription WebSocket messages and connection configuration.
"""
from typing import List, Literal, Optional
from pydantic import BaseModel, Field
class Word(BaseModel):
"""Represents a single word in a transcription with timing and confidence."""
"""Represents a single word in a transcription with timing and confidence.
Parameters:
start: Start time of the word in milliseconds.
end: End time of the word in milliseconds.
text: The transcribed word text.
confidence: Confidence score for the word (0.0 to 1.0).
word_is_final: Whether this word is finalized and won't change.
"""
start: int
end: int
@@ -14,13 +34,23 @@ class Word(BaseModel):
class BaseMessage(BaseModel):
"""Base class for all AssemblyAI WebSocket messages."""
"""Base class for all AssemblyAI WebSocket messages.
Parameters:
type: The message type identifier.
"""
type: str
class BeginMessage(BaseMessage):
"""Message sent when a new session begins."""
"""Message sent when a new session begins.
Parameters:
type: Always "Begin" for this message type.
id: Unique session identifier.
expires_at: Unix timestamp when the session expires.
"""
type: Literal["Begin"] = "Begin"
id: str
@@ -28,7 +58,17 @@ class BeginMessage(BaseMessage):
class TurnMessage(BaseMessage):
"""Message containing transcription data for a turn of speech."""
"""Message containing transcription data for a turn of speech.
Parameters:
type: Always "Turn" for this message type.
turn_order: Sequential number of this turn in the session.
turn_is_formatted: Whether the transcript has been formatted.
end_of_turn: Whether this marks the end of a speaking turn.
transcript: The transcribed text for this turn.
end_of_turn_confidence: Confidence score for end-of-turn detection.
words: List of individual words with timing and confidence data.
"""
type: Literal["Turn"] = "Turn"
turn_order: int
@@ -40,7 +80,13 @@ class TurnMessage(BaseMessage):
class TerminationMessage(BaseMessage):
"""Message sent when the session is terminated."""
"""Message sent when the session is terminated.
Parameters:
type: Always "Termination" for this message type.
audio_duration_seconds: Total duration of audio processed.
session_duration_seconds: Total duration of the session.
"""
type: Literal["Termination"] = "Termination"
audio_duration_seconds: float
@@ -52,6 +98,18 @@ AnyMessage = BeginMessage | TurnMessage | TerminationMessage
class AssemblyAIConnectionParams(BaseModel):
"""Configuration parameters for AssemblyAI WebSocket connection.
Parameters:
sample_rate: Audio sample rate in Hz. Defaults to 16000.
encoding: Audio encoding format. Defaults to "pcm_s16le".
formatted_finals: Whether to enable transcript formatting. Defaults to True.
word_finalization_max_wait_time: Maximum time to wait for word finalization in milliseconds.
end_of_turn_confidence_threshold: Confidence threshold for end-of-turn detection.
min_end_of_turn_silence_when_confident: Minimum silence duration when confident about end-of-turn.
max_turn_silence: Maximum silence duration before forcing end-of-turn.
"""
sample_rate: int = 16000
encoding: Literal["pcm_s16le", "pcm_mulaw"] = "pcm_s16le"
formatted_finals: bool = True

64
src/pipecat/services/assemblyai/stt.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AssemblyAI speech-to-text service implementation.
This module provides integration with AssemblyAI's real-time speech-to-text
WebSocket API for streaming audio transcription.
"""
import asyncio
import json
from typing import Any, AsyncGenerator, Dict
@@ -45,6 +51,13 @@ except ModuleNotFoundError as e:
class AssemblyAISTTService(STTService):
"""AssemblyAI real-time speech-to-text service.
Provides real-time speech transcription using AssemblyAI's WebSocket API.
Supports both interim and final transcriptions with configurable parameters
for audio processing and connection management.
"""
def __init__(
self,
*,
@@ -55,6 +68,16 @@ class AssemblyAISTTService(STTService):
vad_force_turn_endpoint: bool = True,
**kwargs,
):
"""Initialize the AssemblyAI STT service.
Args:
api_key: AssemblyAI API key for authentication.
language: Language code for transcription. Defaults to English (Language.EN).
api_endpoint_base_url: WebSocket endpoint URL. Defaults to AssemblyAI's streaming endpoint.
connection_params: Connection configuration parameters. Defaults to AssemblyAIConnectionParams().
vad_force_turn_endpoint: Whether to force turn endpoint on VAD stop. Defaults to True.
**kwargs: Additional arguments passed to parent STTService class.
"""
self._api_key = api_key
self._language = language
self._api_endpoint_base_url = api_endpoint_base_url
@@ -75,22 +98,50 @@ class AssemblyAISTTService(STTService):
self._chunk_size_bytes = 0
def can_generate_metrics(self) -> bool:
"""Check if the service can generate metrics.
Returns:
True if metrics generation is supported.
"""
return True
async def start(self, frame: StartFrame):
"""Start the speech-to-text service.
Args:
frame: Start frame to begin processing.
"""
await super().start(frame)
self._chunk_size_bytes = int(self._chunk_size_ms * self._sample_rate * 2 / 1000)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the speech-to-text service.
Args:
frame: End frame to stop processing.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the speech-to-text service.
Args:
frame: Cancel frame to abort processing.
"""
await super().cancel(frame)
await self._disconnect()
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Process audio data for speech-to-text conversion.
Args:
audio: Raw audio bytes to process.
Yields:
None (processing handled via WebSocket messages).
"""
self._audio_buffer.extend(audio)
while len(self._audio_buffer) >= self._chunk_size_bytes:
@@ -101,6 +152,12 @@ class AssemblyAISTTService(STTService):
yield None
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames for VAD and metrics handling.
Args:
frame: Frame to process.
direction: Direction of frame processing.
"""
await super().process_frame(frame, direction)
if isinstance(frame, UserStartedSpeakingFrame):
await self.start_ttfb_metrics()
@@ -189,17 +246,16 @@ class AssemblyAISTTService(STTService):
try:
while self._connected:
try:
message = await self._websocket.recv()
self.start_watchdog()
message = await asyncio.wait_for(self._websocket.recv(), timeout=1.0)
data = json.loads(message)
await self._handle_message(data)
except asyncio.TimeoutError:
self.reset_watchdog()
except websockets.exceptions.ConnectionClosedOK:
break
except Exception as e:
logger.error(f"Error processing WebSocket message: {e}")
break
finally:
self.reset_watchdog()
except Exception as e:
logger.error(f"Fatal error in receive handler: {e}")

218
src/pipecat/services/aws/llm.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AWS Bedrock integration for Large Language Model services.
This module provides AWS Bedrock LLM service implementation with support for
Amazon Nova and Anthropic Claude models, including vision capabilities and
function calling.
"""
import asyncio
import base64
import copy
@@ -61,17 +68,44 @@ except ModuleNotFoundError as e:
@dataclass
class AWSBedrockContextAggregatorPair:
"""Container for AWS Bedrock context aggregators.
Provides convenient access to both user and assistant context aggregators
for AWS Bedrock LLM operations.
Parameters:
_user: The user context aggregator instance.
_assistant: The assistant context aggregator instance.
"""
_user: "AWSBedrockUserContextAggregator"
_assistant: "AWSBedrockAssistantContextAggregator"
def user(self) -> "AWSBedrockUserContextAggregator":
"""Get the user context aggregator.
Returns:
The user context aggregator instance.
"""
return self._user
def assistant(self) -> "AWSBedrockAssistantContextAggregator":
"""Get the assistant context aggregator.
Returns:
The assistant context aggregator instance.
"""
return self._assistant
class AWSBedrockLLMContext(OpenAILLMContext):
"""AWS Bedrock-specific LLM context implementation.
Extends OpenAI LLM context to handle AWS Bedrock's specific message format
and system message handling. Manages conversion between OpenAI and Bedrock
message formats.
"""
def __init__(
self,
messages: Optional[List[dict]] = None,
@@ -80,11 +114,27 @@ class AWSBedrockLLMContext(OpenAILLMContext):
*,
system: Optional[str] = None,
):
"""Initialize AWS Bedrock LLM context.
Args:
messages: List of conversation messages in OpenAI format.
tools: List of available function calling tools.
tool_choice: Tool selection strategy or specific tool choice.
system: System message content for AWS Bedrock.
"""
super().__init__(messages=messages, tools=tools, tool_choice=tool_choice)
self.system = system
@staticmethod
def upgrade_to_bedrock(obj: OpenAILLMContext) -> "AWSBedrockLLMContext":
"""Upgrade an OpenAI LLM context to AWS Bedrock format.
Args:
obj: The OpenAI LLM context to upgrade.
Returns:
The upgraded AWS Bedrock LLM context.
"""
logger.debug(f"Upgrading to AWS Bedrock: {obj}")
if isinstance(obj, OpenAILLMContext) and not isinstance(obj, AWSBedrockLLMContext):
obj.__class__ = AWSBedrockLLMContext
@@ -95,6 +145,14 @@ class AWSBedrockLLMContext(OpenAILLMContext):
@classmethod
def from_openai_context(cls, openai_context: OpenAILLMContext):
"""Create AWS Bedrock context from OpenAI context.
Args:
openai_context: The OpenAI LLM context to convert.
Returns:
New AWS Bedrock LLM context instance.
"""
self = cls(
messages=openai_context.messages,
tools=openai_context.tools,
@@ -106,12 +164,28 @@ class AWSBedrockLLMContext(OpenAILLMContext):
@classmethod
def from_messages(cls, messages: List[dict]) -> "AWSBedrockLLMContext":
"""Create AWS Bedrock context from message list.
Args:
messages: List of messages in OpenAI format.
Returns:
New AWS Bedrock LLM context instance.
"""
self = cls(messages=messages)
self._restructure_from_openai_messages()
return self
@classmethod
def from_image_frame(cls, frame: VisionImageRawFrame) -> "AWSBedrockLLMContext":
"""Create AWS Bedrock context from vision image frame.
Args:
frame: The vision image frame to convert.
Returns:
New AWS Bedrock LLM context instance.
"""
context = cls()
context.add_image_frame_message(
format=frame.format, size=frame.size, image=frame.image, text=frame.text
@@ -119,10 +193,14 @@ class AWSBedrockLLMContext(OpenAILLMContext):
return context
def set_messages(self, messages: List):
"""Set the messages list and restructure for Bedrock format.
Args:
messages: List of messages to set.
"""
self._messages[:] = messages
self._restructure_from_openai_messages()
# convert a message in AWS Bedrock format into one or more messages in OpenAI format
def to_standard_messages(self, obj):
"""Convert AWS Bedrock message format to standard structured format.
@@ -295,6 +373,14 @@ class AWSBedrockLLMContext(OpenAILLMContext):
def add_image_frame_message(
self, *, format: str, size: tuple[int, int], image: bytes, text: str = None
):
"""Add an image message to the context.
Args:
format: The image format (e.g., 'RGB', 'RGBA').
size: The image dimensions as (width, height).
image: The raw image data as bytes.
text: Optional text to accompany the image.
"""
buffer = io.BytesIO()
Image.frombytes(format, size, image).save(buffer, format="JPEG")
encoded_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
@@ -306,6 +392,14 @@ class AWSBedrockLLMContext(OpenAILLMContext):
self.add_message({"role": "user", "content": content})
def add_message(self, message):
"""Add a message to the context, merging with previous message if same role.
AWS Bedrock requires alternating roles, so consecutive messages from the
same role are merged together.
Args:
message: The message to add to the context.
"""
try:
if self.messages:
# AWS Bedrock requires that roles alternate. If this message's
@@ -330,10 +424,10 @@ class AWSBedrockLLMContext(OpenAILLMContext):
logger.error(f"Error adding message: {e}")
def _restructure_from_bedrock_messages(self):
"""Restructure messages in AWS Bedrock format by handling system
messages, merging consecutive messages with the same role, and ensuring
proper content formatting.
"""Restructure messages in AWS Bedrock format.
Handles system messages, merging consecutive messages with the same role,
and ensuring proper content formatting.
"""
# Handle system message if present at the beginning
if self.messages and self.messages[0]["role"] == "system":
@@ -416,12 +510,22 @@ class AWSBedrockLLMContext(OpenAILLMContext):
message["content"] = [{"type": "text", "text": "(empty)"}]
def get_messages_for_persistent_storage(self):
"""Get messages formatted for persistent storage.
Returns:
List of messages including system message if present.
"""
messages = super().get_messages_for_persistent_storage()
if self.system:
messages.insert(0, {"role": "system", "content": self.system})
return messages
def get_messages_for_logging(self) -> str:
"""Get messages formatted for logging with sensitive data redacted.
Returns:
JSON string representation of messages with image data redacted.
"""
msgs = []
for message in self.messages:
msg = copy.deepcopy(message)
@@ -435,11 +539,36 @@ class AWSBedrockLLMContext(OpenAILLMContext):
class AWSBedrockUserContextAggregator(LLMUserContextAggregator):
"""User context aggregator for AWS Bedrock LLM service.
Handles aggregation of user messages and frames for AWS Bedrock format.
Inherits all functionality from the base LLM user context aggregator.
Args:
context: The LLM context to aggregate messages into.
params: Configuration parameters for the aggregator.
"""
pass
class AWSBedrockAssistantContextAggregator(LLMAssistantContextAggregator):
"""Assistant context aggregator for AWS Bedrock LLM service.
Handles aggregation of assistant responses and function calls for AWS Bedrock
format, including tool use and tool result handling.
Args:
context: The LLM context to aggregate messages into.
params: Configuration parameters for the aggregator.
"""
async def handle_function_call_in_progress(self, frame: FunctionCallInProgressFrame):
"""Handle function call in progress frame.
Args:
frame: The function call in progress frame to handle.
"""
# Format tool use according to AWS Bedrock API
self._context.add_message(
{
@@ -470,6 +599,11 @@ class AWSBedrockAssistantContextAggregator(LLMAssistantContextAggregator):
)
async def handle_function_call_result(self, frame: FunctionCallResultFrame):
"""Handle function call result frame.
Args:
frame: The function call result frame to handle.
"""
if frame.result:
result = json.dumps(frame.result)
await self._update_function_call_result(frame.function_name, frame.tool_call_id, result)
@@ -479,6 +613,11 @@ class AWSBedrockAssistantContextAggregator(LLMAssistantContextAggregator):
)
async def handle_function_call_cancel(self, frame: FunctionCallCancelFrame):
"""Handle function call cancel frame.
Args:
frame: The function call cancel frame to handle.
"""
await self._update_function_call_result(
frame.function_name, frame.tool_call_id, "CANCELLED"
)
@@ -497,6 +636,11 @@ class AWSBedrockAssistantContextAggregator(LLMAssistantContextAggregator):
content["toolResult"]["content"] = [{"text": result}]
async def handle_user_image_frame(self, frame: UserImageRawFrame):
"""Handle user image frame.
Args:
frame: The user image frame to handle.
"""
await self._update_function_call_result(
frame.request.function_name, frame.request.tool_call_id, "COMPLETED"
)
@@ -509,18 +653,28 @@ class AWSBedrockAssistantContextAggregator(LLMAssistantContextAggregator):
class AWSBedrockLLMService(LLMService):
"""This class implements inference with AWS Bedrock models including Amazon
Nova and Anthropic Claude.
Requires AWS credentials to be configured in the environment or through
boto3 configuration.
"""AWS Bedrock Large Language Model service implementation.
Provides inference capabilities for AWS Bedrock models including Amazon Nova
and Anthropic Claude. Supports streaming responses, function calling, and
vision capabilities.
"""
# Overriding the default adapter to use the Anthropic one.
adapter_class = AWSBedrockLLMAdapter
class InputParams(BaseModel):
"""Input parameters for AWS Bedrock LLM service.
Parameters:
max_tokens: Maximum number of tokens to generate.
temperature: Sampling temperature between 0.0 and 1.0.
top_p: Nucleus sampling parameter between 0.0 and 1.0.
stop_sequences: List of strings that stop generation.
latency: Performance mode - "standard" or "optimized".
additional_model_request_fields: Additional model-specific parameters.
"""
max_tokens: Optional[int] = Field(default_factory=lambda: 4096, ge=1)
temperature: Optional[float] = Field(default_factory=lambda: 0.7, ge=0.0, le=1.0)
top_p: Optional[float] = Field(default_factory=lambda: 0.999, ge=0.0, le=1.0)
@@ -540,6 +694,18 @@ class AWSBedrockLLMService(LLMService):
client_config: Optional[Config] = None,
**kwargs,
):
"""Initialize the AWS Bedrock LLM service.
Args:
model: The AWS Bedrock model identifier to use.
aws_access_key: AWS access key ID. If None, uses default credentials.
aws_secret_key: AWS secret access key. If None, uses default credentials.
aws_session_token: AWS session token for temporary credentials.
aws_region: AWS region for the Bedrock service.
params: Model parameters and configuration.
client_config: Custom boto3 client configuration.
**kwargs: Additional arguments passed to parent LLMService.
"""
super().__init__(**kwargs)
params = params or AWSBedrockLLMService.InputParams()
@@ -573,6 +739,11 @@ class AWSBedrockLLMService(LLMService):
logger.info(f"Using AWS Bedrock model: {model}")
def can_generate_metrics(self) -> bool:
"""Check if the service can generate usage metrics.
Returns:
True if metrics generation is supported.
"""
return True
def create_context_aggregator(
@@ -582,21 +753,21 @@ class AWSBedrockLLMService(LLMService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> AWSBedrockContextAggregatorPair:
"""Create an instance of AWSBedrockContextAggregatorPair from an
OpenAILLMContext. Constructor keyword arguments for both the user and
assistant aggregators can be provided.
"""Create AWS Bedrock-specific context aggregators.
Creates a pair of context aggregators optimized for AWS Bedrocks's message
format, including support for function calls, tool usage, and image handling.
Args:
context (OpenAILLMContext): The LLM context.
user_params (LLMUserAggregatorParams, optional): User aggregator
parameters.
assistant_params (LLMAssistantAggregatorParams, optional): User
aggregator parameters.
context: The LLM context to create aggregators for.
user_params: Parameters for user message aggregation.
assistant_params: Parameters for assistant message aggregation.
Returns:
AWSBedrockContextAggregatorPair: A pair of context aggregators, one
for the user and one for the assistant, encapsulated in an
AWSBedrockContextAggregatorPair: A pair of context aggregators, one for
the user and one for the assistant, encapsulated in an
AWSBedrockContextAggregatorPair.
"""
context.set_llm_adapter(self.get_llm_adapter())
@@ -711,6 +882,8 @@ class AWSBedrockLLMService(LLMService):
function_calls = []
for event in response["stream"]:
self.reset_watchdog()
# Handle text content
if "contentBlockDelta" in event:
delta = event["contentBlockDelta"]["delta"]
@@ -762,6 +935,7 @@ class AWSBedrockLLMService(LLMService):
completion_tokens += usage.get("outputTokens", 0)
cache_read_input_tokens += usage.get("cacheReadInputTokens", 0)
cache_creation_input_tokens += usage.get("cacheWriteInputTokens", 0)
await self.run_function_calls(function_calls)
except asyncio.CancelledError:
# If we're interrupted, we won't get a complete usage report. So set our flag to use the
@@ -789,6 +963,12 @@ class AWSBedrockLLMService(LLMService):
)
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and handle LLM-specific frame types.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
context = None

78
src/pipecat/services/aws/stt.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AWS Transcribe Speech-to-Text service implementation.
This module provides a WebSocket-based connection to AWS Transcribe for real-time
speech-to-text transcription with support for multiple languages and audio formats.
"""
import asyncio
import json
import os
@@ -37,6 +43,13 @@ except ModuleNotFoundError as e:
class AWSTranscribeSTTService(STTService):
"""AWS Transcribe Speech-to-Text service using WebSocket streaming.
Provides real-time speech transcription using AWS Transcribe's streaming API.
Supports multiple languages, configurable sample rates, and both interim and
final transcription results.
"""
def __init__(
self,
*,
@@ -48,6 +61,17 @@ class AWSTranscribeSTTService(STTService):
language: Language = Language.EN,
**kwargs,
):
"""Initialize the AWS Transcribe STT service.
Args:
api_key: AWS secret access key. If None, uses AWS_SECRET_ACCESS_KEY environment variable.
aws_access_key_id: AWS access key ID. If None, uses AWS_ACCESS_KEY_ID environment variable.
aws_session_token: AWS session token for temporary credentials. If None, uses AWS_SESSION_TOKEN environment variable.
region: AWS region for the service. Defaults to "us-east-1".
sample_rate: Audio sample rate in Hz. Must be 8000 or 16000. Defaults to 16000.
language: Language for transcription. Defaults to English.
**kwargs: Additional arguments passed to parent STTService class.
"""
super().__init__(**kwargs)
self._settings = {
@@ -79,14 +103,28 @@ class AWSTranscribeSTTService(STTService):
self._receive_task = None
def get_service_encoding(self, encoding: str) -> str:
"""Convert internal encoding format to AWS Transcribe format."""
"""Convert internal encoding format to AWS Transcribe format.
Args:
encoding: Internal encoding format string.
Returns:
AWS Transcribe compatible encoding format.
"""
encoding_map = {
"linear16": "pcm", # AWS expects "pcm" for 16-bit linear PCM
}
return encoding_map.get(encoding, encoding)
async def start(self, frame: StartFrame):
"""Initialize the connection when the service starts."""
"""Initialize the connection when the service starts.
Args:
frame: Start frame signaling service initialization.
Raises:
RuntimeError: If WebSocket connection cannot be established after retries.
"""
await super().start(frame)
logger.info("Starting AWS Transcribe service...")
retry_count = 0
@@ -108,15 +146,32 @@ class AWSTranscribeSTTService(STTService):
raise RuntimeError("Failed to establish WebSocket connection after multiple attempts")
async def stop(self, frame: EndFrame):
"""Stop the service and disconnect from AWS Transcribe.
Args:
frame: End frame signaling service shutdown.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the service and disconnect from AWS Transcribe.
Args:
frame: Cancel frame signaling service cancellation.
"""
await super().cancel(frame)
await self._disconnect()
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Process audio data and send to AWS Transcribe"""
"""Process audio data and send to AWS Transcribe.
Args:
audio: Raw audio bytes to transcribe.
Yields:
ErrorFrame: If processing fails or connection issues occur.
"""
try:
# Ensure WebSocket is connected
if not self._ws_client or not self._ws_client.open:
@@ -255,7 +310,14 @@ class AWSTranscribeSTTService(STTService):
self._ws_client = None
def language_to_service_language(self, language: Language) -> str | None:
"""Convert internal language enum to AWS Transcribe language code."""
"""Convert internal language enum to AWS Transcribe language code.
Args:
language: Internal language enumeration value.
Returns:
AWS Transcribe compatible language code, or None if unsupported.
"""
language_map = {
Language.EN: "en-US",
Language.ES: "es-US",
@@ -284,9 +346,7 @@ class AWSTranscribeSTTService(STTService):
break
try:
response = await self._ws_client.recv()
self.start_watchdog()
response = await asyncio.wait_for(self._ws_client.recv(), timeout=1.0)
headers, payload = decode_event(response)
@@ -337,6 +397,8 @@ class AWSTranscribeSTTService(STTService):
else:
logger.debug(f"{self} Other message type received: {headers}")
logger.debug(f"{self} Payload: {payload}")
except asyncio.TimeoutError:
self.reset_watchdog()
except websockets.exceptions.ConnectionClosed as e:
logger.error(
f"{self} WebSocket connection closed in receive loop with code {e.code}: {e.reason}"
@@ -345,5 +407,3 @@ class AWSTranscribeSTTService(STTService):
except Exception as e:
logger.error(f"{self} Unexpected error in receive loop: {e}")
break
finally:
self.reset_watchdog()

73
src/pipecat/services/aws/tts.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AWS Polly text-to-speech service implementation.
This module provides integration with Amazon Polly for text-to-speech synthesis,
supporting multiple languages, voices, and SSML features.
"""
import asyncio
import os
from typing import AsyncGenerator, List, Optional
@@ -33,6 +39,14 @@ except ModuleNotFoundError as e:
def language_to_aws_language(language: Language) -> Optional[str]:
"""Convert a Language enum to AWS Polly language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding AWS Polly language code, or None if not supported.
"""
language_map = {
# Arabic
Language.AR: "arb",
@@ -109,7 +123,25 @@ def language_to_aws_language(language: Language) -> Optional[str]:
class AWSPollyTTSService(TTSService):
"""AWS Polly text-to-speech service.
Provides text-to-speech synthesis using Amazon Polly with support for
multiple languages, voices, SSML features, and voice customization
options including prosody controls.
"""
class InputParams(BaseModel):
"""Input parameters for AWS Polly TTS configuration.
Parameters:
engine: TTS engine to use ('standard', 'neural', etc.).
language: Language for synthesis. Defaults to English.
pitch: Voice pitch adjustment (for standard engine only).
rate: Speech rate adjustment.
volume: Voice volume adjustment.
lexicon_names: List of pronunciation lexicons to apply.
"""
engine: Optional[str] = None
language: Optional[Language] = Language.EN
pitch: Optional[str] = None
@@ -129,6 +161,18 @@ class AWSPollyTTSService(TTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initializes the AWS Polly TTS service.
Args:
api_key: AWS secret access key. If None, uses AWS_SECRET_ACCESS_KEY environment variable.
aws_access_key_id: AWS access key ID. If None, uses AWS_ACCESS_KEY_ID environment variable.
aws_session_token: AWS session token for temporary credentials.
region: AWS region for Polly service. Defaults to 'us-east-1'.
voice_id: Voice ID to use for synthesis. Defaults to 'Joanna'.
sample_rate: Audio sample rate. If None, uses service default.
params: Additional input parameters for voice customization.
**kwargs: Additional arguments passed to parent TTSService class.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or AWSPollyTTSService.InputParams()
@@ -174,9 +218,22 @@ class AWSPollyTTSService(TTSService):
)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as AWS Polly service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to AWS Polly language format.
Args:
language: The language to convert.
Returns:
The AWS Polly-specific language code, or None if not supported.
"""
return language_to_aws_language(language)
def _construct_ssml(self, text: str) -> str:
@@ -214,6 +271,15 @@ class AWSPollyTTSService(TTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using AWS Polly.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
def read_audio_data(**args):
response = self._polly_client.synthesize_speech(**args)
if "AudioStream" in response:
@@ -277,7 +343,14 @@ class AWSPollyTTSService(TTSService):
class PollyTTSService(AWSPollyTTSService):
"""Deprecated alias for AWSPollyTTSService."""
def __init__(self, **kwargs):
"""Initialize the deprecated PollyTTSService.
Args:
**kwargs: All arguments passed to AWSPollyTTSService.
"""
super().__init__(**kwargs)
import warnings

107
src/pipecat/services/aws/utils.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AWS Transcribe utility functions and classes for WebSocket streaming.
This module provides utilities for creating presigned URLs, building event messages,
and handling AWS event stream protocol for real-time transcription services.
"""
import binascii
import datetime
import hashlib
@@ -29,7 +35,31 @@ def get_presigned_url(
show_speaker_label: bool = False,
enable_channel_identification: bool = False,
) -> str:
"""Create a presigned URL for AWS Transcribe streaming."""
"""Create a presigned URL for AWS Transcribe streaming.
Args:
region: AWS region for the service.
credentials: Dictionary containing AWS credentials with keys:
- access_key: AWS access key ID
- secret_key: AWS secret access key
- session_token: AWS session token (optional)
language_code: Language code for transcription (e.g., "en-US").
media_encoding: Audio encoding format. Defaults to "pcm".
sample_rate: Audio sample rate in Hz. Defaults to 16000.
number_of_channels: Number of audio channels. Defaults to 1.
enable_partial_results_stabilization: Whether to enable partial result stabilization.
partial_results_stability: Stability level for partial results.
vocabulary_name: Custom vocabulary name to use.
vocabulary_filter_name: Vocabulary filter name to apply.
show_speaker_label: Whether to include speaker labels.
enable_channel_identification: Whether to enable channel identification.
Returns:
Presigned WebSocket URL for AWS Transcribe streaming.
Raises:
ValueError: If required AWS credentials are missing.
"""
access_key = credentials.get("access_key")
secret_key = credentials.get("secret_key")
session_token = credentials.get("session_token")
@@ -58,9 +88,23 @@ def get_presigned_url(
class AWSTranscribePresignedURL:
"""Generator for AWS Transcribe presigned WebSocket URLs.
Handles AWS Signature Version 4 signing process to create authenticated
WebSocket URLs for streaming transcription requests.
"""
def __init__(
self, access_key: str, secret_key: str, session_token: str, region: str = "us-east-1"
):
"""Initialize the presigned URL generator.
Args:
access_key: AWS access key ID.
secret_key: AWS secret access key.
session_token: AWS session token for temporary credentials.
region: AWS region for the service. Defaults to "us-east-1".
"""
self.access_key = access_key
self.secret_key = secret_key
self.session_token = session_token
@@ -96,6 +140,23 @@ class AWSTranscribePresignedURL:
enable_partial_results_stabilization: bool = False,
partial_results_stability: str = "",
) -> str:
"""Generate a presigned WebSocket URL for AWS Transcribe.
Args:
sample_rate: Audio sample rate in Hz.
language_code: Language code for transcription.
media_encoding: Audio encoding format.
vocabulary_name: Custom vocabulary name.
vocabulary_filter_name: Vocabulary filter name.
show_speaker_label: Whether to include speaker labels.
enable_channel_identification: Whether to enable channel identification.
number_of_channels: Number of audio channels.
enable_partial_results_stabilization: Whether to enable partial result stabilization.
partial_results_stability: Stability level for partial results.
Returns:
Presigned WebSocket URL with authentication parameters.
"""
self.endpoint = f"wss://transcribestreaming.{self.region}.amazonaws.com:8443"
self.host = f"transcribestreaming.{self.region}.amazonaws.com:8443"
@@ -172,7 +233,15 @@ class AWSTranscribePresignedURL:
def get_headers(header_name: str, header_value: str) -> bytearray:
"""Build a header following AWS event stream format."""
"""Build a header following AWS event stream format.
Args:
header_name: Name of the header.
header_value: Value of the header.
Returns:
Encoded header as a bytearray following AWS event stream protocol.
"""
name = header_name.encode("utf-8")
name_byte_length = bytes([len(name)])
value_type = bytes([7]) # 7 represents a string
@@ -190,9 +259,21 @@ def get_headers(header_name: str, header_value: str) -> bytearray:
def build_event_message(payload: bytes) -> bytes:
"""
Build an event message for AWS Transcribe streaming.
Matches AWS sample: https://github.com/aws-samples/amazon-transcribe-streaming-python-websockets/blob/main/eventstream.py
"""Build an event message for AWS Transcribe streaming.
Creates a properly formatted AWS event stream message containing audio data
for real-time transcription. Follows the AWS event stream protocol with
prelude, headers, payload, and CRC checksums.
Args:
payload: Raw audio bytes to include in the event message.
Returns:
Complete event message as bytes, ready to send via WebSocket.
Note:
Implementation matches AWS sample:
https://github.com/aws-samples/amazon-transcribe-streaming-python-websockets/blob/main/eventstream.py
"""
# Build headers
content_type_header = get_headers(":content-type", "application/octet-stream")
@@ -235,6 +316,22 @@ def build_event_message(payload: bytes) -> bytes:
def decode_event(message):
"""Decode an AWS event stream message.
Parses an AWS event stream message to extract headers and payload,
verifying CRC checksums for data integrity.
Args:
message: Raw event stream message bytes received from AWS.
Returns:
Tuple containing:
- Dictionary of parsed headers
- Dictionary of parsed JSON payload
Raises:
AssertionError: If CRC checksum verification fails.
"""
# Extract the prelude, headers, payload and CRC
prelude = message[:8]
total_length, headers_length = struct.unpack(">II", prelude)

119
src/pipecat/services/aws_nova_sonic/aws.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""AWS Nova Sonic LLM service implementation for Pipecat AI framework.
This module provides a speech-to-speech LLM service using AWS Nova Sonic, which supports
bidirectional audio streaming, text generation, and function calling capabilities.
"""
import asyncio
import base64
import json
@@ -56,6 +62,7 @@ from pipecat.services.aws_nova_sonic.context import (
)
from pipecat.services.aws_nova_sonic.frames import AWSNovaSonicFunctionCallResultFrame
from pipecat.services.llm_service import LLMService
from pipecat.utils.asyncio.watchdog_coroutine import watchdog_coroutine
from pipecat.utils.time import time_now_iso8601
try:
@@ -83,28 +90,55 @@ except ModuleNotFoundError as e:
class AWSNovaSonicUnhandledFunctionException(Exception):
"""Exception raised when the LLM attempts to call an unregistered function."""
pass
class ContentType(Enum):
"""Content types supported by AWS Nova Sonic.
Parameters:
AUDIO: Audio content type.
TEXT: Text content type.
TOOL: Tool content type.
"""
AUDIO = "AUDIO"
TEXT = "TEXT"
TOOL = "TOOL"
class TextStage(Enum):
"""Text generation stages in AWS Nova Sonic responses.
Parameters:
FINAL: Final text that has been fully generated.
SPECULATIVE: Speculative text that is still being generated.
"""
FINAL = "FINAL" # what has been said
SPECULATIVE = "SPECULATIVE" # what's planned to be said
@dataclass
class CurrentContent:
"""Represents content currently being received from AWS Nova Sonic.
Parameters:
type: The type of content (audio, text, or tool).
role: The role generating the content (user, assistant, etc.).
text_stage: The stage of text generation (final or speculative).
text_content: The actual text content if applicable.
"""
type: ContentType
role: Role
text_stage: TextStage # None if not text
text_content: str # starts as None, then fills in if text
def __str__(self):
"""String representation of the current content."""
return (
f"CurrentContent(\n"
f" type={self.type.name},\n"
@@ -115,6 +149,20 @@ class CurrentContent:
class Params(BaseModel):
"""Configuration parameters for AWS Nova Sonic.
Attributes:
input_sample_rate: Audio input sample rate in Hz.
input_sample_size: Audio input sample size in bits.
input_channel_count: Number of input audio channels.
output_sample_rate: Audio output sample rate in Hz.
output_sample_size: Audio output sample size in bits.
output_channel_count: Number of output audio channels.
max_tokens: Maximum number of tokens to generate.
top_p: Nucleus sampling parameter.
temperature: Sampling temperature for text generation.
"""
# Audio input
input_sample_rate: Optional[int] = Field(default=16000)
input_sample_size: Optional[int] = Field(default=16)
@@ -132,6 +180,12 @@ class Params(BaseModel):
class AWSNovaSonicLLMService(LLMService):
"""AWS Nova Sonic speech-to-speech LLM service.
Provides bidirectional audio streaming, real-time transcription, text generation,
and function calling capabilities using AWS Nova Sonic model.
"""
# Override the default adapter to use the AWSNovaSonicLLMAdapter one
adapter_class = AWSNovaSonicLLMAdapter
@@ -149,6 +203,20 @@ class AWSNovaSonicLLMService(LLMService):
send_transcription_frames: bool = True,
**kwargs,
):
"""Initializes the AWS Nova Sonic LLM service.
Args:
secret_access_key: AWS secret access key for authentication.
access_key_id: AWS access key ID for authentication.
region: AWS region where the service is hosted.
model: Model identifier. Defaults to "amazon.nova-sonic-v1:0".
voice_id: Voice ID for speech synthesis. Options: matthew, tiffany, amy.
params: Model parameters for audio configuration and inference.
system_instruction: System-level instruction for the model.
tools: Available tools/functions for the model to use.
send_transcription_frames: Whether to emit transcription frames.
**kwargs: Additional arguments passed to the parent LLMService.
"""
super().__init__(**kwargs)
self._secret_access_key = secret_access_key
self._access_key_id = access_key_id
@@ -188,16 +256,31 @@ class AWSNovaSonicLLMService(LLMService):
#
async def start(self, frame: StartFrame):
"""Start the service and initiate connection to AWS Nova Sonic.
Args:
frame: The start frame triggering service initialization.
"""
await super().start(frame)
self._wants_connection = True
await self._start_connecting()
async def stop(self, frame: EndFrame):
"""Stop the service and close connections.
Args:
frame: The end frame triggering service shutdown.
"""
await super().stop(frame)
self._wants_connection = False
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the service and close connections.
Args:
frame: The cancel frame triggering service cancellation.
"""
await super().cancel(frame)
self._wants_connection = False
await self._disconnect()
@@ -207,6 +290,11 @@ class AWSNovaSonicLLMService(LLMService):
#
async def reset_conversation(self):
"""Reset the conversation state while preserving context.
Handles bot stopped speaking event, disconnects from the service,
and reconnects with the preserved context.
"""
logger.debug("Resetting conversation")
await self._handle_bot_stopped_speaking(delay_to_catch_trailing_assistant_text=False)
@@ -222,6 +310,12 @@ class AWSNovaSonicLLMService(LLMService):
#
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and handle service-specific logic.
Args:
frame: The frame to process.
direction: The direction the frame is traveling.
"""
await super().process_frame(frame, direction)
if isinstance(frame, OpenAILLMContextFrame):
@@ -697,9 +791,7 @@ class AWSNovaSonicLLMService(LLMService):
try:
while self._stream and not self._disconnecting:
output = await self._stream.await_output()
result = await output[1].receive()
self.start_watchdog()
result = await watchdog_coroutine(output[1].receive(), manager=self.task_manager)
if result.value and result.value.bytes_:
response_data = result.value.bytes_.decode("utf-8")
@@ -728,13 +820,10 @@ class AWSNovaSonicLLMService(LLMService):
elif "completionEnd" in event_json:
# Handle the LLM completion ending
await self._handle_completion_end_event(event_json)
except Exception as e:
logger.error(f"{self} error processing responses: {e}")
if self._wants_connection:
await self.reset_conversation()
finally:
self.reset_watchdog()
async def _handle_completion_start_event(self, event_json):
pass
@@ -961,6 +1050,16 @@ class AWSNovaSonicLLMService(LLMService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> AWSNovaSonicContextAggregatorPair:
"""Create context aggregator pair for managing conversation context.
Args:
context: The OpenAI LLM context to upgrade.
user_params: Parameters for the user context aggregator.
assistant_params: Parameters for the assistant context aggregator.
Returns:
A pair of user and assistant context aggregators.
"""
context.set_llm_adapter(self.get_llm_adapter())
user = AWSNovaSonicUserContextAggregator(context=context, params=user_params)
@@ -979,6 +1078,14 @@ class AWSNovaSonicLLMService(LLMService):
)
async def trigger_assistant_response(self):
"""Trigger an assistant response by sending audio cue.
Sends a pre-recorded "ready" audio trigger to prompt the assistant
to start speaking. This is useful for controlling conversation flow.
Returns:
False if already triggering a response, True otherwise.
"""
if self._triggering_assistant_response:
return False

140
src/pipecat/services/aws_nova_sonic/context.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Context management for AWS Nova Sonic LLM service.
This module provides specialized context aggregators and message handling for AWS Nova Sonic,
including conversation history management and role-specific message processing.
"""
import copy
from dataclasses import dataclass, field
from enum import Enum
@@ -35,6 +41,15 @@ from pipecat.services.openai.llm import (
class Role(Enum):
"""Roles supported in AWS Nova Sonic conversations.
Parameters:
SYSTEM: System-level messages (not used in conversation history).
USER: Messages sent by the user.
ASSISTANT: Messages sent by the assistant.
TOOL: Messages sent by tools (not used in conversation history).
"""
SYSTEM = "SYSTEM"
USER = "USER"
ASSISTANT = "ASSISTANT"
@@ -43,18 +58,45 @@ class Role(Enum):
@dataclass
class AWSNovaSonicConversationHistoryMessage:
"""A single message in AWS Nova Sonic conversation history.
Parameters:
role: The role of the message sender (USER or ASSISTANT only).
text: The text content of the message.
"""
role: Role # only USER and ASSISTANT
text: str
@dataclass
class AWSNovaSonicConversationHistory:
"""Complete conversation history for AWS Nova Sonic initialization.
Parameters:
system_instruction: System-level instruction for the conversation.
messages: List of conversation messages between user and assistant.
"""
system_instruction: str = None
messages: list[AWSNovaSonicConversationHistoryMessage] = field(default_factory=list)
class AWSNovaSonicLLMContext(OpenAILLMContext):
"""Specialized LLM context for AWS Nova Sonic service.
Extends OpenAI context with Nova Sonic-specific message handling,
conversation history management, and text buffering capabilities.
"""
def __init__(self, messages=None, tools=None, **kwargs):
"""Initialize AWS Nova Sonic LLM context.
Args:
messages: Initial messages for the context.
tools: Available tools for the context.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(messages=messages, tools=tools, **kwargs)
self.__setup_local()
@@ -67,6 +109,15 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
def upgrade_to_nova_sonic(
obj: OpenAILLMContext, system_instruction: str
) -> "AWSNovaSonicLLMContext":
"""Upgrade an OpenAI context to AWS Nova Sonic context.
Args:
obj: The OpenAI context to upgrade.
system_instruction: System instruction for the context.
Returns:
The upgraded AWS Nova Sonic context.
"""
if isinstance(obj, OpenAILLMContext) and not isinstance(obj, AWSNovaSonicLLMContext):
obj.__class__ = AWSNovaSonicLLMContext
obj.__setup_local(system_instruction)
@@ -74,6 +125,14 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
# NOTE: this method has the side-effect of updating _system_instruction from messages
def get_messages_for_initializing_history(self) -> AWSNovaSonicConversationHistory:
"""Get conversation history for initializing AWS Nova Sonic session.
Processes stored messages and extracts system instruction and conversation
history in the format expected by AWS Nova Sonic.
Returns:
Formatted conversation history with system instruction and messages.
"""
history = AWSNovaSonicConversationHistory(system_instruction=self._system_instruction)
# Bail if there are no messages
@@ -103,6 +162,11 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
return history
def get_messages_for_persistent_storage(self):
"""Get messages formatted for persistent storage.
Returns:
List of messages including system instruction if present.
"""
messages = super().get_messages_for_persistent_storage()
# If we have a system instruction and messages doesn't already contain it, add it
if self._system_instruction and not (messages and messages[0].get("role") == "system"):
@@ -110,6 +174,14 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
return messages
def from_standard_message(self, message) -> AWSNovaSonicConversationHistoryMessage:
"""Convert standard message format to Nova Sonic format.
Args:
message: Standard message dictionary to convert.
Returns:
Nova Sonic conversation history message, or None if not convertible.
"""
role = message.get("role")
if message.get("role") == "user" or message.get("role") == "assistant":
content = message.get("content")
@@ -131,10 +203,20 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
# Sonic conversation history
def buffer_user_text(self, text):
"""Buffer user text for later flushing to context.
Args:
text: User text to buffer.
"""
self._user_text += f" {text}" if self._user_text else text
# logger.debug(f"User text buffered: {self._user_text}")
def flush_aggregated_user_text(self) -> str:
"""Flush buffered user text to context as a complete message.
Returns:
The flushed user text, or empty string if no text was buffered.
"""
if not self._user_text:
return ""
user_text = self._user_text
@@ -148,10 +230,16 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
return user_text
def buffer_assistant_text(self, text):
"""Buffer assistant text for later flushing to context.
Args:
text: Assistant text to buffer.
"""
self._assistant_text += text
# logger.debug(f"Assistant text buffered: {self._assistant_text}")
def flush_aggregated_assistant_text(self):
"""Flush buffered assistant text to context as a complete message."""
if not self._assistant_text:
return
message = {
@@ -165,13 +253,31 @@ class AWSNovaSonicLLMContext(OpenAILLMContext):
@dataclass
class AWSNovaSonicMessagesUpdateFrame(DataFrame):
"""Frame containing updated AWS Nova Sonic context.
Parameters:
context: The updated AWS Nova Sonic LLM context.
"""
context: AWSNovaSonicLLMContext
class AWSNovaSonicUserContextAggregator(OpenAIUserContextAggregator):
"""Context aggregator for user messages in AWS Nova Sonic conversations.
Extends the OpenAI user context aggregator to emit Nova Sonic-specific
context update frames.
"""
async def process_frame(
self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM
):
"""Process frames and emit Nova Sonic-specific context updates.
Args:
frame: The frame to process.
direction: The direction the frame is traveling.
"""
await super().process_frame(frame, direction)
# Parent does not push LLMMessagesUpdateFrame
@@ -180,7 +286,19 @@ class AWSNovaSonicUserContextAggregator(OpenAIUserContextAggregator):
class AWSNovaSonicAssistantContextAggregator(OpenAIAssistantContextAggregator):
"""Context aggregator for assistant messages in AWS Nova Sonic conversations.
Provides specialized handling for assistant responses and function calls
in AWS Nova Sonic context, with custom frame processing logic.
"""
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames with Nova Sonic-specific logic.
Args:
frame: The frame to process.
direction: The direction the frame is traveling.
"""
# HACK: For now, disable the context aggregator by making it just pass through all frames
# that the parent handles (except the function call stuff, which we still need).
# For an explanation of this hack, see
@@ -205,6 +323,11 @@ class AWSNovaSonicAssistantContextAggregator(OpenAIAssistantContextAggregator):
await super().process_frame(frame, direction)
async def handle_function_call_result(self, frame: FunctionCallResultFrame):
"""Handle function call results for AWS Nova Sonic.
Args:
frame: The function call result frame to handle.
"""
await super().handle_function_call_result(frame)
# The standard function callback code path pushes the FunctionCallResultFrame from the LLM
@@ -217,11 +340,28 @@ class AWSNovaSonicAssistantContextAggregator(OpenAIAssistantContextAggregator):
@dataclass
class AWSNovaSonicContextAggregatorPair:
"""Pair of user and assistant context aggregators for AWS Nova Sonic.
Parameters:
_user: The user context aggregator.
_assistant: The assistant context aggregator.
"""
_user: AWSNovaSonicUserContextAggregator
_assistant: AWSNovaSonicAssistantContextAggregator
def user(self) -> AWSNovaSonicUserContextAggregator:
"""Get the user context aggregator.
Returns:
The user context aggregator instance.
"""
return self._user
def assistant(self) -> AWSNovaSonicAssistantContextAggregator:
"""Get the assistant context aggregator.
Returns:
The assistant context aggregator instance.
"""
return self._assistant

11
src/pipecat/services/aws_nova_sonic/frames.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Custom frames for AWS Nova Sonic LLM service."""
from dataclasses import dataclass
from pipecat.frames.frames import DataFrame, FunctionCallResultFrame
@@ -11,4 +13,13 @@ from pipecat.frames.frames import DataFrame, FunctionCallResultFrame
@dataclass
class AWSNovaSonicFunctionCallResultFrame(DataFrame):
"""Frame containing function call result for AWS Nova Sonic processing.
This frame wraps a standard function call result frame to enable
AWS Nova Sonic-specific handling and context updates.
Parameters:
result_frame: The underlying function call result frame.
"""
result_frame: FunctionCallResultFrame

12
src/pipecat/services/azure/common.py
View File

@@ -4,14 +4,22 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
from typing import Optional
"""Language conversion utilities for Azure services."""
from loguru import logger
from typing import Optional
from pipecat.transcriptions.language import Language
def language_to_azure_language(language: Language) -> Optional[str]:
"""Convert a Language enum to Azure language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding Azure language code, or None if not supported.
"""
language_map = {
# Afrikaans
Language.AF: "af-ZA",

32
src/pipecat/services/azure/image.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Azure OpenAI image generation service implementation.
This module provides integration with Azure's OpenAI image generation API
using REST endpoints for creating images from text prompts.
"""
import asyncio
import io
from typing import AsyncGenerator
@@ -17,6 +23,13 @@ from pipecat.services.image_service import ImageGenService
class AzureImageGenServiceREST(ImageGenService):
"""Azure OpenAI REST-based image generation service.
Provides image generation using Azure's OpenAI service via REST API.
Supports asynchronous image generation with polling for completion
and automatic image download and processing.
"""
def __init__(
self,
*,
@@ -27,6 +40,16 @@ class AzureImageGenServiceREST(ImageGenService):
aiohttp_session: aiohttp.ClientSession,
api_version="2023-06-01-preview",
):
"""Initialize the AzureImageGenServiceREST.
Args:
image_size: Size specification for generated images (e.g., "1024x1024").
api_key: Azure OpenAI API key for authentication.
endpoint: Azure OpenAI endpoint URL.
model: The image generation model to use.
aiohttp_session: Shared aiohttp session for HTTP requests.
api_version: Azure API version string. Defaults to "2023-06-01-preview".
"""
super().__init__()
self._api_key = api_key
@@ -37,6 +60,15 @@ class AzureImageGenServiceREST(ImageGenService):
self._aiohttp_session = aiohttp_session
async def run_image_gen(self, prompt: str) -> AsyncGenerator[Frame, None]:
"""Generate an image from a text prompt using Azure OpenAI.
Args:
prompt: The text prompt describing the desired image.
Yields:
URLImageRawFrame containing the generated image data, or
ErrorFrame if generation fails.
"""
url = f"{self._azure_endpoint}openai/images/generations:submit?api-version={self._api_version}"
headers = {"api-key": self._api_key, "Content-Type": "application/json"}

29
src/pipecat/services/azure/llm.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Azure OpenAI service implementation for the Pipecat AI framework."""
from loguru import logger
from openai import AsyncAzureOpenAI
@@ -15,13 +17,6 @@ class AzureLLMService(OpenAILLMService):
This service extends OpenAILLMService to connect to Azure's OpenAI endpoint while
maintaining full compatibility with OpenAI's interface and functionality.
Args:
api_key (str): The API key for accessing Azure OpenAI
endpoint (str): The Azure endpoint URL
model (str): The model identifier to use
api_version (str, optional): Azure API version. Defaults to "2024-09-01-preview"
**kwargs: Additional keyword arguments passed to OpenAILLMService
"""
def __init__(
@@ -33,6 +28,15 @@ class AzureLLMService(OpenAILLMService):
api_version: str = "2024-09-01-preview",
**kwargs,
):
"""Initialize the Azure LLM service.
Args:
api_key: The API key for accessing Azure OpenAI.
endpoint: The Azure endpoint URL.
model: The model identifier to use.
api_version: Azure API version. Defaults to "2024-09-01-preview".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
# Initialize variables before calling parent __init__() because that
# will call create_client() and we need those values there.
self._endpoint = endpoint
@@ -40,7 +44,16 @@ class AzureLLMService(OpenAILLMService):
super().__init__(api_key=api_key, model=model, **kwargs)
def create_client(self, api_key=None, base_url=None, **kwargs):
"""Create OpenAI-compatible client for Azure OpenAI endpoint."""
"""Create OpenAI-compatible client for Azure OpenAI endpoint.
Args:
api_key: API key for authentication. Uses instance key if None.
base_url: Base URL for the client. Ignored for Azure implementation.
**kwargs: Additional keyword arguments. Ignored for Azure implementation.
Returns:
AsyncAzureOpenAI: Configured Azure OpenAI client instance.
"""
logger.debug(f"Creating Azure OpenAI client with endpoint {self._endpoint}")
return AsyncAzureOpenAI(
api_key=api_key,

60
src/pipecat/services/azure/stt.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Azure Speech-to-Text service implementation for Pipecat.
This module provides speech-to-text functionality using Azure Cognitive Services
Speech SDK for real-time audio transcription.
"""
import asyncio
from typing import AsyncGenerator, Optional
@@ -40,6 +46,13 @@ except ModuleNotFoundError as e:
class AzureSTTService(STTService):
"""Azure Speech-to-Text service for real-time audio transcription.
This service uses Azure Cognitive Services Speech SDK to convert speech
audio into text transcriptions. It supports continuous recognition and
provides real-time transcription results with timing information.
"""
def __init__(
self,
*,
@@ -49,6 +62,15 @@ class AzureSTTService(STTService):
sample_rate: Optional[int] = None,
**kwargs,
):
"""Initialize the Azure STT service.
Args:
api_key: Azure Cognitive Services subscription key.
region: Azure region for the Speech service (e.g., 'eastus').
language: Language for speech recognition. Defaults to English (US).
sample_rate: Audio sample rate in Hz. If None, uses service default.
**kwargs: Additional arguments passed to parent STTService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
self._speech_config = SpeechConfig(
@@ -66,9 +88,25 @@ class AzureSTTService(STTService):
}
def can_generate_metrics(self) -> bool:
"""Check if this service can generate performance metrics.
Returns:
True as this service supports metrics generation.
"""
return True
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Process audio data for speech-to-text conversion.
Feeds audio data to the Azure speech recognizer for processing.
Recognition results are handled asynchronously through callbacks.
Args:
audio: Raw audio bytes to process.
Yields:
None - actual transcription frames are pushed via callbacks.
"""
await self.start_processing_metrics()
await self.start_ttfb_metrics()
if self._audio_stream:
@@ -76,6 +114,14 @@ class AzureSTTService(STTService):
yield None
async def start(self, frame: StartFrame):
"""Start the speech recognition service.
Initializes the Azure speech recognizer with audio stream configuration
and begins continuous speech recognition.
Args:
frame: Frame indicating the start of processing.
"""
await super().start(frame)
if self._audio_stream:
@@ -93,6 +139,13 @@ class AzureSTTService(STTService):
self._speech_recognizer.start_continuous_recognition_async()
async def stop(self, frame: EndFrame):
"""Stop the speech recognition service.
Cleanly shuts down the Azure speech recognizer and closes audio streams.
Args:
frame: Frame indicating the end of processing.
"""
await super().stop(frame)
if self._speech_recognizer:
@@ -102,6 +155,13 @@ class AzureSTTService(STTService):
self._audio_stream.close()
async def cancel(self, frame: CancelFrame):
"""Cancel the speech recognition service.
Immediately stops recognition and closes resources.
Args:
frame: Frame indicating cancellation.
"""
await super().cancel(frame)
if self._speech_recognizer:

110
src/pipecat/services/azure/tts.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Azure Cognitive Services Text-to-Speech service implementations."""
import asyncio
from typing import AsyncGenerator, Optional
@@ -39,6 +41,15 @@ except ModuleNotFoundError as e:
def sample_rate_to_output_format(sample_rate: int) -> SpeechSynthesisOutputFormat:
"""Convert sample rate to Azure speech synthesis output format.
Args:
sample_rate: Sample rate in Hz.
Returns:
Corresponding Azure SpeechSynthesisOutputFormat enum value.
Defaults to Raw24Khz16BitMonoPcm if sample rate not found.
"""
sample_rate_map = {
8000: SpeechSynthesisOutputFormat.Raw8Khz16BitMonoPcm,
16000: SpeechSynthesisOutputFormat.Raw16Khz16BitMonoPcm,
@@ -51,7 +62,26 @@ def sample_rate_to_output_format(sample_rate: int) -> SpeechSynthesisOutputForma
class AzureBaseTTSService(TTSService):
"""Base class for Azure Cognitive Services text-to-speech implementations.
Provides common functionality for Azure TTS services including SSML
construction, voice configuration, and parameter management.
"""
class InputParams(BaseModel):
"""Input parameters for Azure TTS voice configuration.
Parameters:
emphasis: Emphasis level for speech ("strong", "moderate", "reduced").
language: Language for synthesis. Defaults to English (US).
pitch: Voice pitch adjustment (e.g., "+10%", "-5Hz", "high").
rate: Speech rate multiplier. Defaults to "1.05".
role: Voice role for expression (e.g., "YoungAdultFemale").
style: Speaking style (e.g., "cheerful", "sad", "excited").
style_degree: Intensity of the speaking style (0.01 to 2.0).
volume: Volume level (e.g., "+20%", "loud", "x-soft").
"""
emphasis: Optional[str] = None
language: Optional[Language] = Language.EN_US
pitch: Optional[str] = None
@@ -71,6 +101,16 @@ class AzureBaseTTSService(TTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the Azure TTS service with configuration parameters.
Args:
api_key: Azure Cognitive Services subscription key.
region: Azure region identifier (e.g., "eastus", "westus2").
voice: Voice name to use for synthesis. Defaults to "en-US-SaraNeural".
sample_rate: Audio sample rate in Hz. If None, uses service default.
params: Voice and synthesis parameters configuration.
**kwargs: Additional arguments passed to parent TTSService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or AzureBaseTTSService.InputParams()
@@ -94,9 +134,22 @@ class AzureBaseTTSService(TTSService):
self._speech_synthesizer = None
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Azure TTS service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Azure language format.
Args:
language: The language to convert.
Returns:
The Azure-specific language code, or None if not supported.
"""
return language_to_azure_language(language)
def _construct_ssml(self, text: str) -> str:
@@ -146,13 +199,30 @@ class AzureBaseTTSService(TTSService):
class AzureTTSService(AzureBaseTTSService):
"""Azure Cognitive Services streaming TTS service.
Provides real-time text-to-speech synthesis using Azure's WebSocket-based
streaming API. Audio chunks are streamed as they become available for
lower latency playback.
"""
def __init__(self, **kwargs):
"""Initialize the Azure streaming TTS service.
Args:
**kwargs: All arguments passed to AzureBaseTTSService parent class.
"""
super().__init__(**kwargs)
self._speech_config = None
self._speech_synthesizer = None
self._audio_queue = asyncio.Queue()
async def start(self, frame: StartFrame):
"""Start the Azure TTS service and initialize speech synthesizer.
Args:
frame: Start frame containing initialization parameters.
"""
await super().start(frame)
if self._speech_config:
@@ -183,24 +253,33 @@ class AzureTTSService(AzureBaseTTSService):
self._speech_synthesizer.synthesis_canceled.connect(self._handle_canceled)
def _handle_synthesizing(self, evt):
"""Handle audio chunks as they arrive"""
"""Handle audio chunks as they arriv."""
if evt.result and evt.result.audio_data:
self._audio_queue.put_nowait(evt.result.audio_data)
def _handle_completed(self, evt):
"""Handle synthesis completion"""
"""Handle synthesis completion."""
self._audio_queue.put_nowait(None) # Signal completion
def _handle_canceled(self, evt):
"""Handle synthesis cancellation"""
"""Handle synthesis cancellation."""
logger.error(f"Speech synthesis canceled: {evt.result.cancellation_details.reason}")
self._audio_queue.put_nowait(None)
async def flush_audio(self):
"""Flush any pending audio data."""
logger.trace(f"{self}: flushing audio")
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Azure's streaming synthesis.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing synthesized speech data.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:
@@ -244,12 +323,29 @@ class AzureTTSService(AzureBaseTTSService):
class AzureHttpTTSService(AzureBaseTTSService):
"""Azure Cognitive Services HTTP-based TTS service.
Provides text-to-speech synthesis using Azure's HTTP API for simpler,
non-streaming synthesis. Suitable for use cases where streaming is not
required and simpler integration is preferred.
"""
def __init__(self, **kwargs):
"""Initialize the Azure HTTP TTS service.
Args:
**kwargs: All arguments passed to AzureBaseTTSService parent class.
"""
super().__init__(**kwargs)
self._speech_config = None
self._speech_synthesizer = None
async def start(self, frame: StartFrame):
"""Start the Azure HTTP TTS service and initialize speech synthesizer.
Args:
frame: Start frame containing initialization parameters.
"""
await super().start(frame)
if self._speech_config:
@@ -269,6 +365,14 @@ class AzureHttpTTSService(AzureBaseTTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Azure's HTTP synthesis API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the complete synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
await self.start_ttfb_metrics()

99
src/pipecat/services/cartesia/stt.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Cartesia Speech-to-Text service implementation.
This module provides a WebSocket-based STT service that integrates with
the Cartesia Live transcription API for real-time speech recognition.
"""
import asyncio
import json
import urllib.parse
@@ -30,6 +36,12 @@ from pipecat.utils.tracing.service_decorators import traced_stt
class CartesiaLiveOptions:
"""Configuration options for Cartesia Live STT service.
Manages transcription parameters including model selection, language,
audio encoding format, and sample rate settings.
"""
def __init__(
self,
*,
@@ -39,6 +51,15 @@ class CartesiaLiveOptions:
sample_rate: int = 16000,
**kwargs,
):
"""Initialize CartesiaLiveOptions with default or provided parameters.
Args:
model: The transcription model to use. Defaults to "ink-whisper".
language: Target language for transcription. Defaults to English.
encoding: Audio encoding format. Defaults to "pcm_s16le".
sample_rate: Audio sample rate in Hz. Defaults to 16000.
**kwargs: Additional parameters for the transcription service.
"""
self.model = model
self.language = language
self.encoding = encoding
@@ -46,6 +67,11 @@ class CartesiaLiveOptions:
self.additional_params = kwargs
def to_dict(self):
"""Convert options to dictionary format.
Returns:
Dictionary containing all configuration parameters.
"""
params = {
"model": self.model,
"language": self.language if isinstance(self.language, str) else self.language.value,
@@ -56,19 +82,48 @@ class CartesiaLiveOptions:
return params
def items(self):
"""Get configuration items as key-value pairs.
Returns:
Iterator of (key, value) tuples for all configuration parameters.
"""
return self.to_dict().items()
def get(self, key, default=None):
"""Get a configuration value by key.
Args:
key: The configuration parameter name to retrieve.
default: Default value if key is not found.
Returns:
The configuration value or default if not found.
"""
if hasattr(self, key):
return getattr(self, key)
return self.additional_params.get(key, default)
@classmethod
def from_json(cls, json_str: str) -> "CartesiaLiveOptions":
"""Create options from JSON string.
Args:
json_str: JSON string containing configuration parameters.
Returns:
New CartesiaLiveOptions instance with parsed parameters.
"""
return cls(**json.loads(json_str))
class CartesiaSTTService(STTService):
"""Speech-to-text service using Cartesia Live API.
Provides real-time speech transcription through WebSocket connection
to Cartesia's Live transcription service. Supports both interim and
final transcriptions with configurable models and languages.
"""
def __init__(
self,
*,
@@ -78,6 +133,15 @@ class CartesiaSTTService(STTService):
live_options: Optional[CartesiaLiveOptions] = None,
**kwargs,
):
"""Initialize CartesiaSTTService with API key and options.
Args:
api_key: Authentication key for Cartesia API.
base_url: Custom API endpoint URL. If empty, uses default.
sample_rate: Audio sample rate in Hz. Defaults to 16000.
live_options: Configuration options for transcription service.
**kwargs: Additional arguments passed to parent STTService.
"""
sample_rate = sample_rate or (live_options.sample_rate if live_options else None)
super().__init__(sample_rate=sample_rate, **kwargs)
@@ -108,21 +172,49 @@ class CartesiaSTTService(STTService):
self._receiver_task = None
def can_generate_metrics(self) -> bool:
"""Check if the service can generate processing metrics.
Returns:
True, indicating metrics are supported.
"""
return True
async def start(self, frame: StartFrame):
"""Start the STT service and establish connection.
Args:
frame: Frame indicating service should start.
"""
await super().start(frame)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the STT service and close connection.
Args:
frame: Frame indicating service should stop.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the STT service and close connection.
Args:
frame: Frame indicating service should be cancelled.
"""
await super().cancel(frame)
await self._disconnect()
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Process audio data for speech-to-text transcription.
Args:
audio: Raw audio bytes to transcribe.
Yields:
None - transcription results are handled via WebSocket responses.
"""
# If the connection is closed, due to timeout, we need to reconnect when the user starts speaking again
if not self._connection or self._connection.closed:
await self._connect()
@@ -225,10 +317,17 @@ class CartesiaSTTService(STTService):
self._connection = None
async def start_metrics(self):
"""Start performance metrics collection for transcription processing."""
await self.start_ttfb_metrics()
await self.start_processing_metrics()
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and handle speech events.
Args:
frame: The frame to process.
direction: Direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)
if isinstance(frame, UserStartedSpeakingFrame):

152
src/pipecat/services/cartesia/tts.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Cartesia text-to-speech service implementations."""
import base64
import json
import uuid
@@ -27,6 +29,7 @@ from pipecat.frames.frames import (
from pipecat.processors.frame_processor import FrameDirection
from pipecat.services.tts_service import AudioContextWordTTSService, TTSService
from pipecat.transcriptions.language import Language
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.text.base_text_aggregator import BaseTextAggregator
from pipecat.utils.text.skip_tags_aggregator import SkipTagsAggregator
from pipecat.utils.tracing.service_decorators import traced_tts
@@ -42,6 +45,14 @@ except ModuleNotFoundError as e:
def language_to_cartesia_language(language: Language) -> Optional[str]:
"""Convert a Language enum to Cartesia language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding Cartesia language code, or None if not supported.
"""
BASE_LANGUAGES = {
Language.DE: "de",
Language.EN: "en",
@@ -74,7 +85,22 @@ def language_to_cartesia_language(language: Language) -> Optional[str]:
class CartesiaTTSService(AudioContextWordTTSService):
"""Cartesia TTS service with WebSocket streaming and word timestamps.
Provides text-to-speech using Cartesia's streaming WebSocket API.
Supports word-level timestamps, audio context management, and various voice
customization options including speed and emotion controls.
"""
class InputParams(BaseModel):
"""Input parameters for Cartesia TTS configuration.
Parameters:
language: Language to use for synthesis.
speed: Voice speed control (string or float).
emotion: List of emotion controls (deprecated).
"""
language: Optional[Language] = Language.EN
speed: Optional[Union[str, float]] = ""
emotion: Optional[List[str]] = []
@@ -94,6 +120,21 @@ class CartesiaTTSService(AudioContextWordTTSService):
text_aggregator: Optional[BaseTextAggregator] = None,
**kwargs,
):
"""Initialize the Cartesia TTS service.
Args:
api_key: Cartesia API key for authentication.
voice_id: ID of the voice to use for synthesis.
cartesia_version: API version string for Cartesia service.
url: WebSocket URL for Cartesia TTS API.
model: TTS model to use (e.g., "sonic-2").
sample_rate: Audio sample rate. If None, uses default.
encoding: Audio encoding format.
container: Audio container format.
params: Additional input parameters for voice customization.
text_aggregator: Custom text aggregator for processing input text.
**kwargs: Additional arguments passed to the parent service.
"""
# Aggregating sentences still gives cleaner-sounding results and fewer
# artifacts than streaming one word at a time. On average, waiting for a
# full sentence should only "cost" us 15ms or so with GPT-4o or a Llama
@@ -137,14 +178,32 @@ class CartesiaTTSService(AudioContextWordTTSService):
self._receive_task = None
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Cartesia service supports metrics generation.
"""
return True
async def set_model(self, model: str):
"""Set the TTS model.
Args:
model: The model name to use for synthesis.
"""
self._model_id = model
await super().set_model(model)
logger.info(f"Switching TTS model to: [{model}]")
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Cartesia language format.
Args:
language: The language to convert.
Returns:
The Cartesia-specific language code, or None if not supported.
"""
return language_to_cartesia_language(language)
def _build_msg(
@@ -182,15 +241,30 @@ class CartesiaTTSService(AudioContextWordTTSService):
return json.dumps(msg)
async def start(self, frame: StartFrame):
"""Start the Cartesia TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._settings["output_format"]["sample_rate"] = self.sample_rate
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the Cartesia TTS service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Stop the Cartesia TTS service.
Args:
frame: The end frame.
"""
await super().cancel(frame)
await self._disconnect()
@@ -247,6 +321,7 @@ class CartesiaTTSService(AudioContextWordTTSService):
self._context_id = None
async def flush_audio(self):
"""Flush any pending audio and finalize the current context."""
if not self._context_id or not self._websocket:
return
logger.trace(f"{self}: flushing audio")
@@ -255,7 +330,9 @@ class CartesiaTTSService(AudioContextWordTTSService):
self._context_id = None
async def _receive_messages(self):
async for message in self._get_websocket():
async for message in WatchdogAsyncIterator(
self._get_websocket(), manager=self.task_manager
):
msg = json.loads(message)
if not msg or not self.audio_context_available(msg["context_id"]):
continue
@@ -287,6 +364,14 @@ class CartesiaTTSService(AudioContextWordTTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Cartesia's streaming API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:
@@ -316,7 +401,22 @@ class CartesiaTTSService(AudioContextWordTTSService):
class CartesiaHttpTTSService(TTSService):
"""Cartesia HTTP-based TTS service.
Provides text-to-speech using Cartesia's HTTP API for simpler, non-streaming
synthesis. Suitable for use cases where streaming is not required and simpler
integration is preferred.
"""
class InputParams(BaseModel):
"""Input parameters for Cartesia HTTP TTS configuration.
Parameters:
language: Language to use for synthesis.
speed: Voice speed control (string or float).
emotion: List of emotion controls (deprecated).
"""
language: Optional[Language] = Language.EN
speed: Optional[Union[str, float]] = ""
emotion: Optional[List[str]] = Field(default_factory=list)
@@ -335,6 +435,20 @@ class CartesiaHttpTTSService(TTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the Cartesia HTTP TTS service.
Args:
api_key: Cartesia API key for authentication.
voice_id: ID of the voice to use for synthesis.
model: TTS model to use (e.g., "sonic-2").
base_url: Base URL for Cartesia HTTP API.
cartesia_version: API version string for Cartesia service.
sample_rate: Audio sample rate. If None, uses default.
encoding: Audio encoding format.
container: Audio container format.
params: Additional input parameters for voice customization.
**kwargs: Additional arguments passed to the parent TTSService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or CartesiaHttpTTSService.InputParams()
@@ -363,25 +477,61 @@ class CartesiaHttpTTSService(TTSService):
)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Cartesia HTTP service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Cartesia language format.
Args:
language: The language to convert.
Returns:
The Cartesia-specific language code, or None if not supported.
"""
return language_to_cartesia_language(language)
async def start(self, frame: StartFrame):
"""Start the Cartesia HTTP TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._settings["output_format"]["sample_rate"] = self.sample_rate
async def stop(self, frame: EndFrame):
"""Stop the Cartesia HTTP TTS service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._client.close()
async def cancel(self, frame: CancelFrame):
"""Cancel the Cartesia HTTP TTS service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._client.close()
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Cartesia's HTTP API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:

39
src/pipecat/services/cerebras/llm.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Cerebras LLM service implementation using OpenAI-compatible interface."""
from typing import List
from loguru import logger
@@ -19,12 +21,6 @@ class CerebrasLLMService(OpenAILLMService):
This service extends OpenAILLMService to connect to Cerebras's API endpoint while
maintaining full compatibility with OpenAI's interface and functionality.
Args:
api_key (str): The API key for accessing Cerebras's API
base_url (str, optional): The base URL for Cerebras API. Defaults to "https://api.cerebras.ai/v1"
model (str, optional): The model identifier to use. Defaults to "llama-3.3-70b"
**kwargs: Additional keyword arguments passed to OpenAILLMService
"""
def __init__(
@@ -35,10 +31,27 @@ class CerebrasLLMService(OpenAILLMService):
model: str = "llama-3.3-70b",
**kwargs,
):
"""Initialize the Cerebras LLM service.
Args:
api_key: The API key for accessing Cerebras's API.
base_url: The base URL for Cerebras API. Defaults to "https://api.cerebras.ai/v1".
model: The model identifier to use. Defaults to "llama-3.3-70b".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
def create_client(self, api_key=None, base_url=None, **kwargs):
"""Create OpenAI-compatible client for Cerebras API endpoint."""
"""Create OpenAI-compatible client for Cerebras API endpoint.
Args:
api_key: The API key for authentication. If None, uses instance key.
base_url: The base URL for the API. If None, uses instance URL.
**kwargs: Additional arguments passed to the client constructor.
Returns:
An OpenAI-compatible client configured for Cerebras API.
"""
logger.debug(f"Creating Cerebras client with api {base_url}")
return super().create_client(api_key, base_url, **kwargs)
@@ -48,14 +61,14 @@ class CerebrasLLMService(OpenAILLMService):
"""Create a streaming chat completion using Cerebras's API.
Args:
context (OpenAILLMContext): The context object containing tools configuration
and other settings for the chat completion.
messages (List[ChatCompletionMessageParam]): The list of messages comprising
the conversation history and current request.
context: The context object containing tools configuration
and other settings for the chat completion.
messages: The list of messages comprising
the conversation history and current request.
Returns:
AsyncStream[ChatCompletionChunk]: A streaming response of chat completion
chunks that can be processed asynchronously.
A streaming response of chat completion
chunks that can be processed asynchronously.
"""
params = {
"model": self.model_name,

70
src/pipecat/services/deepgram/stt.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Deepgram speech-to-text service implementation."""
from typing import AsyncGenerator, Dict, Optional
from loguru import logger
@@ -41,6 +43,13 @@ except ModuleNotFoundError as e:
class DeepgramSTTService(STTService):
"""Deepgram speech-to-text service.
Provides real-time speech recognition using Deepgram's WebSocket API.
Supports configurable models, languages, VAD events, and various audio
processing options.
"""
def __init__(
self,
*,
@@ -52,6 +61,17 @@ class DeepgramSTTService(STTService):
addons: Optional[Dict] = None,
**kwargs,
):
"""Initialize the Deepgram STT service.
Args:
api_key: Deepgram API key for authentication.
url: Deprecated. Use base_url instead.
base_url: Custom Deepgram API base URL.
sample_rate: Audio sample rate. If None, uses default or live_options value.
live_options: Deepgram LiveOptions for detailed configuration.
addons: Additional Deepgram features to enable.
**kwargs: Additional arguments passed to the parent STTService.
"""
sample_rate = sample_rate or (live_options.sample_rate if live_options else None)
super().__init__(sample_rate=sample_rate, **kwargs)
@@ -108,12 +128,27 @@ class DeepgramSTTService(STTService):
@property
def vad_enabled(self):
"""Check if Deepgram VAD events are enabled.
Returns:
True if VAD events are enabled in the current settings.
"""
return self._settings["vad_events"]
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Deepgram service supports metrics generation.
"""
return True
async def set_model(self, model: str):
"""Set the Deepgram model and reconnect.
Args:
model: The Deepgram model name to use.
"""
await super().set_model(model)
logger.info(f"Switching STT model to: [{model}]")
self._settings["model"] = model
@@ -121,25 +156,53 @@ class DeepgramSTTService(STTService):
await self._connect()
async def set_language(self, language: Language):
"""Set the recognition language and reconnect.
Args:
language: The language to use for speech recognition.
"""
logger.info(f"Switching STT language to: [{language}]")
self._settings["language"] = language
await self._disconnect()
await self._connect()
async def start(self, frame: StartFrame):
"""Start the Deepgram STT service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._settings["sample_rate"] = self.sample_rate
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the Deepgram STT service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the Deepgram STT service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._disconnect()
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Send audio data to Deepgram for transcription.
Args:
audio: Raw audio bytes to transcribe.
Yields:
Frame: None (transcription results come via WebSocket callbacks).
"""
await self._connection.send(audio)
yield None
@@ -172,6 +235,7 @@ class DeepgramSTTService(STTService):
await self._connection.finish()
async def start_metrics(self):
"""Start TTFB and processing metrics collection."""
await self.start_ttfb_metrics()
await self.start_processing_metrics()
@@ -235,6 +299,12 @@ class DeepgramSTTService(STTService):
)
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames with Deepgram-specific handling.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
if isinstance(frame, UserStartedSpeakingFrame) and not self.vad_enabled:

36
src/pipecat/services/deepgram/tts.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Deepgram text-to-speech service implementation.
This module provides integration with Deepgram's text-to-speech API
for generating speech from text using various voice models.
"""
from typing import AsyncGenerator, Optional
from loguru import logger
@@ -27,6 +33,13 @@ except ModuleNotFoundError as e:
class DeepgramTTSService(TTSService):
"""Deepgram text-to-speech service.
Provides text-to-speech synthesis using Deepgram's streaming API.
Supports various voice models and audio encoding formats with
configurable sample rates and quality settings.
"""
def __init__(
self,
*,
@@ -37,6 +50,16 @@ class DeepgramTTSService(TTSService):
encoding: str = "linear16",
**kwargs,
):
"""Initialize the Deepgram TTS service.
Args:
api_key: Deepgram API key for authentication.
voice: Voice model to use for synthesis. Defaults to "aura-2-helena-en".
base_url: Custom base URL for Deepgram API. Uses default if empty.
sample_rate: Audio sample rate in Hz. If None, uses service default.
encoding: Audio encoding format. Defaults to "linear16".
**kwargs: Additional arguments passed to parent TTSService class.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
self._settings = {
@@ -48,10 +71,23 @@ class DeepgramTTSService(TTSService):
self._deepgram_client = DeepgramClient(api_key, config=client_options)
def can_generate_metrics(self) -> bool:
"""Check if the service can generate metrics.
Returns:
True, as Deepgram TTS service supports metrics generation.
"""
return True
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Deepgram's TTS API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech, plus start/stop frames.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
options = SpeakOptions(

40
src/pipecat/services/deepseek/llm.py
View File

@@ -4,6 +4,7 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""DeepSeek LLM service implementation using OpenAI-compatible interface."""
from typing import List
@@ -20,12 +21,6 @@ class DeepSeekLLMService(OpenAILLMService):
This service extends OpenAILLMService to connect to DeepSeek's API endpoint while
maintaining full compatibility with OpenAI's interface and functionality.
Args:
api_key (str): The API key for accessing DeepSeek's API
base_url (str, optional): The base URL for DeepSeek API. Defaults to "https://api.deepseek.com/v1"
model (str, optional): The model identifier to use. Defaults to "deepseek-chat"
**kwargs: Additional keyword arguments passed to OpenAILLMService
"""
def __init__(
@@ -36,27 +31,44 @@ class DeepSeekLLMService(OpenAILLMService):
model: str = "deepseek-chat",
**kwargs,
):
"""Initialize the DeepSeek LLM service.
Args:
api_key: The API key for accessing DeepSeek's API.
base_url: The base URL for DeepSeek API. Defaults to "https://api.deepseek.com/v1".
model: The model identifier to use. Defaults to "deepseek-chat".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
def create_client(self, api_key=None, base_url=None, **kwargs):
"""Create OpenAI-compatible client for DeepSeek API endpoint."""
"""Create OpenAI-compatible client for DeepSeek API endpoint.
Args:
api_key: The API key for authentication. If None, uses instance default.
base_url: The base URL for the API. If None, uses instance default.
**kwargs: Additional keyword arguments for client configuration.
Returns:
An OpenAI-compatible client configured for DeepSeek's API.
"""
logger.debug(f"Creating DeepSeek client with api {base_url}")
return super().create_client(api_key, base_url, **kwargs)
async def get_chat_completions(
self, context: OpenAILLMContext, messages: List[ChatCompletionMessageParam]
) -> AsyncStream[ChatCompletionChunk]:
"""Create a streaming chat completion using Cerebras's API.
"""Create a streaming chat completion using DeepSeek's API.
Args:
context (OpenAILLMContext): The context object containing tools configuration
and other settings for the chat completion.
messages (List[ChatCompletionMessageParam]): The list of messages comprising
the conversation history and current request.
context: The context object containing tools configuration
and other settings for the chat completion.
messages: The list of messages comprising the conversation
history and current request.
Returns:
AsyncStream[ChatCompletionChunk]: A streaming response of chat completion
chunks that can be processed asynchronously.
A streaming response of chat completion chunks that can be
processed asynchronously.
"""
params = {
"model": self.model_name,

201
src/pipecat/services/elevenlabs/tts.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""ElevenLabs text-to-speech service implementations.
This module provides WebSocket and HTTP-based TTS services using ElevenLabs API
with support for streaming audio, word timestamps, and voice customization.
"""
import asyncio
import base64
import json
@@ -32,6 +38,7 @@ from pipecat.services.tts_service import (
WordTTSService,
)
from pipecat.transcriptions.language import Language
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.tracing.service_decorators import traced_tts
# See .env.example for ElevenLabs configuration needed
@@ -56,6 +63,14 @@ ELEVENLABS_MULTILINGUAL_MODELS = {
def language_to_elevenlabs_language(language: Language) -> Optional[str]:
"""Convert a Language enum to ElevenLabs language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding ElevenLabs language code, or None if not supported.
"""
BASE_LANGUAGES = {
Language.AR: "ar",
Language.BG: "bg",
@@ -105,6 +120,14 @@ def language_to_elevenlabs_language(language: Language) -> Optional[str]:
def output_format_from_sample_rate(sample_rate: int) -> str:
"""Get the appropriate output format string for a given sample rate.
Args:
sample_rate: The audio sample rate in Hz.
Returns:
The ElevenLabs output format string.
"""
match sample_rate:
case 8000:
return "pcm_8000"
@@ -128,10 +151,10 @@ def build_elevenlabs_voice_settings(
"""Build voice settings dictionary for ElevenLabs based on provided settings.
Args:
settings: Dictionary containing voice settings parameters
settings: Dictionary containing voice settings parameters.
Returns:
Dictionary of voice settings or None if no valid settings are provided
Dictionary of voice settings or None if no valid settings are provided.
"""
voice_setting_keys = ["stability", "similarity_boost", "style", "use_speaker_boost", "speed"]
@@ -146,6 +169,15 @@ def build_elevenlabs_voice_settings(
def calculate_word_times(
alignment_info: Mapping[str, Any], cumulative_time: float
) -> List[Tuple[str, float]]:
"""Calculate word timestamps from character alignment information.
Args:
alignment_info: Character alignment data from ElevenLabs API.
cumulative_time: Base time offset for this chunk.
Returns:
List of (word, timestamp) tuples.
"""
zipped_times = list(zip(alignment_info["chars"], alignment_info["charStartTimesMs"]))
words = "".join(alignment_info["chars"]).split(" ")
@@ -165,7 +197,28 @@ def calculate_word_times(
class ElevenLabsTTSService(AudioContextWordTTSService):
"""ElevenLabs WebSocket-based TTS service with word timestamps.
Provides real-time text-to-speech using ElevenLabs' WebSocket streaming API.
Supports word-level timestamps, audio context management, and various voice
customization options including stability, similarity boost, and speed controls.
"""
class InputParams(BaseModel):
"""Input parameters for ElevenLabs TTS configuration.
Parameters:
language: Language to use for synthesis.
stability: Voice stability control (0.0 to 1.0).
similarity_boost: Similarity boost control (0.0 to 1.0).
style: Style control for voice expression (0.0 to 1.0).
use_speaker_boost: Whether to use speaker boost enhancement.
speed: Voice speed control (0.25 to 4.0).
auto_mode: Whether to enable automatic mode optimization.
enable_ssml_parsing: Whether to parse SSML tags in text.
enable_logging: Whether to enable ElevenLabs logging.
"""
language: Optional[Language] = None
stability: Optional[float] = None
similarity_boost: Optional[float] = None
@@ -187,6 +240,17 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the ElevenLabs TTS service.
Args:
api_key: ElevenLabs API key for authentication.
voice_id: ID of the voice to use for synthesis.
model: TTS model to use (e.g., "eleven_flash_v2_5").
url: WebSocket URL for ElevenLabs TTS API.
sample_rate: Audio sample rate. If None, uses default.
params: Additional input parameters for voice customization.
**kwargs: Additional arguments passed to the parent service.
"""
# Aggregating sentences still gives cleaner-sounding results and fewer
# artifacts than streaming one word at a time. On average, waiting for a
# full sentence should only "cost" us 15ms or so with GPT-4o or a Llama
@@ -243,21 +307,40 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
self._keepalive_task = None
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as ElevenLabs service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to ElevenLabs language format.
Args:
language: The language to convert.
Returns:
The ElevenLabs-specific language code, or None if not supported.
"""
return language_to_elevenlabs_language(language)
def _set_voice_settings(self):
return build_elevenlabs_voice_settings(self._settings)
async def set_model(self, model: str):
"""Set the TTS model and reconnect.
Args:
model: The model name to use for synthesis.
"""
await super().set_model(model)
logger.info(f"Switching TTS model to: [{model}]")
await self._disconnect()
await self._connect()
async def _update_settings(self, settings: Mapping[str, Any]):
"""Update service settings and reconnect if voice changed."""
prev_voice = self._voice_id
await super()._update_settings(settings)
if not prev_voice == self._voice_id:
@@ -266,19 +349,35 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
await self._connect()
async def start(self, frame: StartFrame):
"""Start the ElevenLabs TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._output_format = output_format_from_sample_rate(self.sample_rate)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the ElevenLabs TTS service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the ElevenLabs TTS service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._disconnect()
async def flush_audio(self):
"""Flush any pending audio and finalize the current context."""
if not self._context_id or not self._websocket:
return
logger.trace(f"{self}: flushing audio")
@@ -286,6 +385,12 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
await self._websocket.send(json.dumps(msg))
async def push_frame(self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM):
"""Push a frame and handle state changes.
Args:
frame: The frame to push.
direction: The direction to push the frame.
"""
await super().push_frame(frame, direction)
if isinstance(frame, (TTSStoppedFrame, StartInterruptionFrame)):
self._started = False
@@ -373,6 +478,7 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
raise Exception("Websocket not connected")
async def _handle_interruption(self, frame: StartInterruptionFrame, direction: FrameDirection):
"""Handle interruption by closing the current context."""
await super()._handle_interruption(frame, direction)
# Close the current context when interrupted without closing the websocket
@@ -394,7 +500,10 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
self._started = False
async def _receive_messages(self):
async for message in self._get_websocket():
"""Handle incoming WebSocket messages from ElevenLabs."""
async for message in WatchdogAsyncIterator(
self._get_websocket(), manager=self.task_manager
):
msg = json.loads(message)
received_ctx_id = msg.get("contextId")
@@ -425,8 +534,11 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
self._cumulative_time = word_times[-1][1]
async def _keepalive_task_handler(self):
"""Send periodic keepalive messages to maintain WebSocket connection."""
KEEPALIVE_SLEEP = 10 if self.task_manager.task_watchdog_enabled else 3
while True:
await asyncio.sleep(10)
self.reset_watchdog()
await asyncio.sleep(KEEPALIVE_SLEEP)
try:
if self._websocket and self._websocket.open:
if self._context_id:
@@ -448,12 +560,21 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
break
async def _send_text(self, text: str):
"""Send text to the WebSocket for synthesis."""
if self._websocket and self._context_id:
msg = {"text": text, "context_id": self._context_id}
await self._websocket.send(json.dumps(msg))
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using ElevenLabs' streaming WebSocket API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:
@@ -492,19 +613,26 @@ class ElevenLabsTTSService(AudioContextWordTTSService):
class ElevenLabsHttpTTSService(WordTTSService):
"""ElevenLabs Text-to-Speech service using HTTP streaming with word timestamps.
"""ElevenLabs HTTP-based TTS service with word timestamps.
Args:
api_key: ElevenLabs API key
voice_id: ID of the voice to use
aiohttp_session: aiohttp ClientSession
model: Model ID (default: "eleven_flash_v2_5" for low latency)
base_url: API base URL
sample_rate: Output sample rate
params: Additional parameters for voice configuration
Provides text-to-speech using ElevenLabs' HTTP streaming API for simpler,
non-WebSocket integration. Suitable for use cases where streaming WebSocket
connection is not required or desired.
"""
class InputParams(BaseModel):
"""Input parameters for ElevenLabs HTTP TTS configuration.
Parameters:
language: Language to use for synthesis.
optimize_streaming_latency: Latency optimization level (0-4).
stability: Voice stability control (0.0 to 1.0).
similarity_boost: Similarity boost control (0.0 to 1.0).
style: Style control for voice expression (0.0 to 1.0).
use_speaker_boost: Whether to use speaker boost enhancement.
speed: Voice speed control (0.25 to 4.0).
"""
language: Optional[Language] = None
optimize_streaming_latency: Optional[int] = None
stability: Optional[float] = None
@@ -525,6 +653,18 @@ class ElevenLabsHttpTTSService(WordTTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the ElevenLabs HTTP TTS service.
Args:
api_key: ElevenLabs API key for authentication.
voice_id: ID of the voice to use for synthesis.
aiohttp_session: aiohttp ClientSession for HTTP requests.
model: TTS model to use (e.g., "eleven_flash_v2_5").
base_url: Base URL for ElevenLabs HTTP API.
sample_rate: Audio sample rate. If None, uses default.
params: Additional input parameters for voice customization.
**kwargs: Additional arguments passed to the parent service.
"""
super().__init__(
aggregate_sentences=True,
push_text_frames=False,
@@ -564,11 +704,22 @@ class ElevenLabsHttpTTSService(WordTTSService):
self._previous_text = ""
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert pipecat Language to ElevenLabs language code."""
"""Convert pipecat Language to ElevenLabs language code.
Args:
language: The language to convert.
Returns:
The ElevenLabs-specific language code, or None if not supported.
"""
return language_to_elevenlabs_language(language)
def can_generate_metrics(self) -> bool:
"""Indicate that this service can generate usage metrics."""
"""Check if this service can generate processing metrics.
Returns:
True, as ElevenLabs HTTP service supports metrics generation.
"""
return True
def _set_voice_settings(self):
@@ -582,12 +733,22 @@ class ElevenLabsHttpTTSService(WordTTSService):
logger.debug(f"{self}: Reset internal state")
async def start(self, frame: StartFrame):
"""Initialize the service upon receiving a StartFrame."""
"""Start the ElevenLabs HTTP TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._output_format = output_format_from_sample_rate(self.sample_rate)
self._reset_state()
async def push_frame(self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM):
"""Push a frame and handle state changes.
Args:
frame: The frame to push.
direction: The direction to push the frame.
"""
await super().push_frame(frame, direction)
if isinstance(frame, (StartInterruptionFrame, TTSStoppedFrame)):
# Reset timing on interruption or stop
@@ -614,10 +775,10 @@ class ElevenLabsHttpTTSService(WordTTSService):
[("Hello", 0.1), ("world", 0.5)]
Args:
alignment_info: Character timing data from ElevenLabs
alignment_info: Character timing data from ElevenLabs.
Returns:
List of (word, timestamp) pairs
List of (word, timestamp) pairs.
"""
chars = alignment_info.get("characters", [])
char_start_times = alignment_info.get("character_start_times_seconds", [])
@@ -668,10 +829,10 @@ class ElevenLabsHttpTTSService(WordTTSService):
Includes previous text as context for better prosody continuity.
Args:
text: Text to convert to speech
text: Text to convert to speech.
Yields:
Audio and control frames
Frame: Audio and control frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")

43
src/pipecat/services/fal/image.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Fal's image generation service implementation.
This module provides integration with Fal's image generation API
for creating images from text prompts using various AI models.
"""
import asyncio
import io
import os
@@ -26,7 +32,25 @@ except ModuleNotFoundError as e:
class FalImageGenService(ImageGenService):
"""Fal's image generation service.
Provides text-to-image generation using Fal.ai's API with configurable
parameters for image quality, safety, and format options.
"""
class InputParams(BaseModel):
"""Input parameters for Fal.ai image generation.
Parameters:
seed: Random seed for reproducible generation. If None, uses random seed.
num_inference_steps: Number of inference steps for generation. Defaults to 8.
num_images: Number of images to generate. Defaults to 1.
image_size: Image dimensions as string preset or dict with width/height. Defaults to "square_hd".
expand_prompt: Whether to automatically expand/enhance the prompt. Defaults to False.
enable_safety_checker: Whether to enable content safety filtering. Defaults to True.
format: Output image format. Defaults to "png".
"""
seed: Optional[int] = None
num_inference_steps: int = 8
num_images: int = 1
@@ -44,6 +68,15 @@ class FalImageGenService(ImageGenService):
key: Optional[str] = None,
**kwargs,
):
"""Initialize the FalImageGenService.
Args:
params: Input parameters for image generation configuration.
aiohttp_session: HTTP client session for downloading generated images.
model: The Fal.ai model to use for generation. Defaults to "fal-ai/fast-sdxl".
key: Optional API key for Fal.ai. If provided, sets FAL_KEY environment variable.
**kwargs: Additional arguments passed to parent ImageGenService.
"""
super().__init__(**kwargs)
self.set_model_name(model)
self._params = params
@@ -52,6 +85,16 @@ class FalImageGenService(ImageGenService):
os.environ["FAL_KEY"] = key
async def run_image_gen(self, prompt: str) -> AsyncGenerator[Frame, None]:
"""Generate an image from a text prompt.
Args:
prompt: The text prompt to generate an image from.
Yields:
URLImageRawFrame: Frame containing the generated image data and metadata.
ErrorFrame: If image generation fails.
"""
def load_image_bytes(encoded_image: bytes):
buffer = io.BytesIO(encoded_image)
image = Image.open(buffer)

56
src/pipecat/services/fal/stt.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Fal speech-to-text service implementation.
This module provides integration with Fal's Wizper API for speech-to-text
transcription using segmented audio processing.
"""
import os
from typing import AsyncGenerator, Optional
@@ -27,7 +33,14 @@ except ModuleNotFoundError as e:
def language_to_fal_language(language: Language) -> Optional[str]:
"""Language support for Fal's Wizper API."""
"""Convert a Language enum to Fal's Wizper language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding Fal Wizper language code, or None if not supported.
"""
BASE_LANGUAGES = {
Language.AF: "af",
Language.AM: "am",
@@ -145,18 +158,12 @@ class FalSTTService(SegmentedSTTService):
This service uses Fal's Wizper API to perform speech-to-text transcription on audio
segments. It inherits from SegmentedSTTService to handle audio buffering and speech detection.
Args:
api_key: Fal API key. If not provided, will check FAL_KEY environment variable.
sample_rate: Audio sample rate in Hz. If not provided, uses the pipeline's rate.
params: Configuration parameters for the Wizper API.
**kwargs: Additional arguments passed to SegmentedSTTService.
"""
class InputParams(BaseModel):
"""Configuration parameters for Fal's Wizper API.
Attributes:
Parameters:
language: Language of the audio input. Defaults to English.
task: Task to perform ('transcribe' or 'translate'). Defaults to 'transcribe'.
chunk_level: Level of chunking ('segment'). Defaults to 'segment'.
@@ -176,6 +183,14 @@ class FalSTTService(SegmentedSTTService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the FalSTTService with API key and parameters.
Args:
api_key: Fal API key. If not provided, will check FAL_KEY environment variable.
sample_rate: Audio sample rate in Hz. If not provided, uses the pipeline's rate.
params: Configuration parameters for the Wizper API.
**kwargs: Additional arguments passed to SegmentedSTTService.
"""
super().__init__(
sample_rate=sample_rate,
**kwargs,
@@ -201,16 +216,39 @@ class FalSTTService(SegmentedSTTService):
}
def can_generate_metrics(self) -> bool:
"""Check if the service can generate processing metrics.
Returns:
True, as Fal STT service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Fal's service-specific language code.
Args:
language: The language to convert.
Returns:
The Fal-specific language code, or None if not supported.
"""
return language_to_fal_language(language)
async def set_language(self, language: Language):
"""Set the transcription language.
Args:
language: The language to use for speech-to-text transcription.
"""
logger.info(f"Switching STT language to: [{language}]")
self._settings["language"] = self.language_to_service_language(language)
async def set_model(self, model: str):
"""Set the STT model.
Args:
model: The model name to use for transcription.
"""
await super().set_model(model)
logger.info(f"Switching STT model to: [{model}]")
@@ -229,7 +267,7 @@ class FalSTTService(SegmentedSTTService):
audio: Raw audio bytes in WAV format (already converted by base class).
Yields:
Frame: TranscriptionFrame containing the transcribed text.
Frame: TranscriptionFrame containing the transcribed text, or ErrorFrame on failure.
Note:
The audio is already in WAV format from the SegmentedSTTService.

36
src/pipecat/services/fireworks/llm.py
View File

@@ -4,6 +4,7 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Fireworks AI service implementation using OpenAI-compatible interface."""
from typing import List
@@ -19,12 +20,6 @@ class FireworksLLMService(OpenAILLMService):
This service extends OpenAILLMService to connect to Fireworks' API endpoint while
maintaining full compatibility with OpenAI's interface and functionality.
Args:
api_key (str): The API key for accessing Fireworks AI
model (str, optional): The model identifier to use. Defaults to "accounts/fireworks/models/firefunction-v2"
base_url (str, optional): The base URL for Fireworks API. Defaults to "https://api.fireworks.ai/inference/v1"
**kwargs: Additional keyword arguments passed to OpenAILLMService
"""
def __init__(
@@ -35,10 +30,27 @@ class FireworksLLMService(OpenAILLMService):
base_url: str = "https://api.fireworks.ai/inference/v1",
**kwargs,
):
"""Initialize the Fireworks LLM service.
Args:
api_key: The API key for accessing Fireworks AI.
model: The model identifier to use. Defaults to "accounts/fireworks/models/firefunction-v2".
base_url: The base URL for Fireworks API. Defaults to "https://api.fireworks.ai/inference/v1".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
def create_client(self, api_key=None, base_url=None, **kwargs):
"""Create OpenAI-compatible client for Fireworks API endpoint."""
"""Create OpenAI-compatible client for Fireworks API endpoint.
Args:
api_key: API key for authentication. If None, uses instance default.
base_url: Base URL for the API. If None, uses instance default.
**kwargs: Additional arguments passed to the client constructor.
Returns:
Configured OpenAI client instance for Fireworks API.
"""
logger.debug(f"Creating Fireworks client with api {base_url}")
return super().create_client(api_key, base_url, **kwargs)
@@ -47,7 +59,15 @@ class FireworksLLMService(OpenAILLMService):
):
"""Get chat completions from Fireworks API.
Removes OpenAI-specific parameters not supported by Fireworks.
Removes OpenAI-specific parameters not supported by Fireworks and
configures the request with Fireworks-compatible settings.
Args:
context: The OpenAI LLM context containing tools and settings.
messages: List of chat completion message parameters.
Returns:
Async generator yielding chat completion chunks from Fireworks API.
"""
params = {
"model": self.model_name,

65
src/pipecat/services/fish/tts.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Fish Audio text-to-speech service implementation.
This module provides integration with Fish Audio's real-time TTS WebSocket API
for streaming text-to-speech synthesis with customizable voice parameters.
"""
import uuid
from typing import AsyncGenerator, Literal, Optional
@@ -39,7 +45,23 @@ FishAudioOutputFormat = Literal["opus", "mp3", "pcm", "wav"]
class FishAudioTTSService(InterruptibleTTSService):
"""Fish Audio text-to-speech service with WebSocket streaming.
Provides real-time text-to-speech synthesis using Fish Audio's WebSocket API.
Supports various audio formats, customizable prosody controls, and streaming
audio generation with interruption handling.
"""
class InputParams(BaseModel):
"""Input parameters for Fish Audio TTS configuration.
Parameters:
language: Language for synthesis. Defaults to English.
latency: Latency mode ("normal" or "balanced"). Defaults to "normal".
prosody_speed: Speech speed multiplier (0.5-2.0). Defaults to 1.0.
prosody_volume: Volume adjustment in dB. Defaults to 0.
"""
language: Optional[Language] = Language.EN
latency: Optional[str] = "normal" # "normal" or "balanced"
prosody_speed: Optional[float] = 1.0 # Speech speed (0.5-2.0)
@@ -55,6 +77,16 @@ class FishAudioTTSService(InterruptibleTTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the Fish Audio TTS service.
Args:
api_key: Fish Audio API key for authentication.
model: Reference ID of the voice model to use for synthesis.
output_format: Audio output format. Defaults to "pcm".
sample_rate: Audio sample rate. If None, uses default.
params: Additional input parameters for voice customization.
**kwargs: Additional arguments passed to the parent service.
"""
super().__init__(
push_stop_frames=True,
pause_frame_processing=True,
@@ -85,23 +117,48 @@ class FishAudioTTSService(InterruptibleTTSService):
self.set_model_name(model)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Fish Audio service supports metrics generation.
"""
return True
async def set_model(self, model: str):
"""Set the TTS model (reference ID).
Args:
model: The reference ID of the voice model to use.
"""
self._settings["reference_id"] = model
await super().set_model(model)
logger.info(f"Switching TTS model to: [{model}]")
async def start(self, frame: StartFrame):
"""Start the Fish Audio TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._settings["sample_rate"] = self.sample_rate
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the Fish Audio TTS service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the Fish Audio TTS service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._disconnect()
@@ -191,6 +248,14 @@ class FishAudioTTSService(InterruptibleTTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Fish Audio's streaming API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames and control frames for the synthesized speech.
"""
logger.debug(f"{self}: Generating Fish TTS: [{text}]")
try:
if not self._websocket or self._websocket.closed:

228
src/pipecat/services/gemini_multimodal_live/events.py
View File

@@ -3,7 +3,8 @@
#
# SPDX-License-Identifier: BSD 2-Clause License
#
#
"""Event models and utilities for Google Gemini Multimodal Live API."""
import base64
import io
@@ -23,11 +24,25 @@ from pipecat.frames.frames import ImageRawFrame
class MediaChunk(BaseModel):
"""Represents a chunk of media data for transmission.
Parameters:
mimeType: MIME type of the media content.
data: Base64-encoded media data.
"""
mimeType: str
data: str
class ContentPart(BaseModel):
"""Represents a part of content that can contain text or media.
Parameters:
text: Text content. Defaults to None.
inlineData: Inline media data. Defaults to None.
"""
text: Optional[str] = Field(default=None, validate_default=False)
inlineData: Optional[MediaChunk] = Field(default=None, validate_default=False)
fileData: Optional['FileData'] = Field(default=None, validate_default=False)
@@ -43,6 +58,13 @@ ContentPart.model_rebuild() # Rebuild model to resolve forward reference
class Turn(BaseModel):
"""Represents a conversational turn in the dialogue.
Parameters:
role: The role of the speaker, either "user" or "model". Defaults to "user".
parts: List of content parts that make up the turn.
"""
role: Literal["user", "model"] = "user"
parts: List[ContentPart]
@@ -64,7 +86,15 @@ class EndSensitivity(str, Enum):
class AutomaticActivityDetection(BaseModel):
"""Configures automatic detection of activity."""
"""Configures automatic detection of voice activity.
Parameters:
disabled: Whether automatic activity detection is disabled. Defaults to None.
start_of_speech_sensitivity: Sensitivity for detecting speech start. Defaults to None.
prefix_padding_ms: Padding before speech start in milliseconds. Defaults to None.
end_of_speech_sensitivity: Sensitivity for detecting speech end. Defaults to None.
silence_duration_ms: Duration of silence to detect speech end. Defaults to None.
"""
disabled: Optional[bool] = None
start_of_speech_sensitivity: Optional[StartSensitivity] = None
@@ -74,25 +104,57 @@ class AutomaticActivityDetection(BaseModel):
class RealtimeInputConfig(BaseModel):
"""Configures the realtime input behavior."""
"""Configures the realtime input behavior.
Parameters:
automatic_activity_detection: Voice activity detection configuration. Defaults to None.
"""
automatic_activity_detection: Optional[AutomaticActivityDetection] = None
class RealtimeInput(BaseModel):
"""Contains realtime input media chunks.
Parameters:
mediaChunks: List of media chunks for realtime processing.
"""
mediaChunks: List[MediaChunk]
class ClientContent(BaseModel):
"""Content sent from client to the Gemini Live API.
Parameters:
turns: List of conversation turns. Defaults to None.
turnComplete: Whether the client's turn is complete. Defaults to False.
"""
turns: Optional[List[Turn]] = None
turnComplete: bool = False
class AudioInputMessage(BaseModel):
"""Message containing audio input data.
Parameters:
realtimeInput: Realtime input containing audio chunks.
"""
realtimeInput: RealtimeInput
@classmethod
def from_raw_audio(cls, raw_audio: bytes, sample_rate: int) -> "AudioInputMessage":
"""Create an audio input message from raw audio data.
Args:
raw_audio: Raw audio bytes.
sample_rate: Audio sample rate in Hz.
Returns:
AudioInputMessage instance with encoded audio data.
"""
data = base64.b64encode(raw_audio).decode("utf-8")
return cls(
realtimeInput=RealtimeInput(
@@ -102,10 +164,24 @@ class AudioInputMessage(BaseModel):
class VideoInputMessage(BaseModel):
"""Message containing video/image input data.
Parameters:
realtimeInput: Realtime input containing video/image chunks.
"""
realtimeInput: RealtimeInput
@classmethod
def from_image_frame(cls, frame: ImageRawFrame) -> "VideoInputMessage":
"""Create a video input message from an image frame.
Args:
frame: Image frame to encode.
Returns:
VideoInputMessage instance with encoded image data.
"""
buffer = io.BytesIO()
Image.frombytes(frame.format, frame.size, frame.image).save(buffer, format="JPEG")
data = base64.b64encode(buffer.getvalue()).decode("utf-8")
@@ -115,18 +191,44 @@ class VideoInputMessage(BaseModel):
class ClientContentMessage(BaseModel):
"""Message containing client content for the API.
Parameters:
clientContent: The client content to send.
"""
clientContent: ClientContent
class SystemInstruction(BaseModel):
"""System instruction for the model.
Parameters:
parts: List of content parts that make up the system instruction.
"""
parts: List[ContentPart]
class AudioTranscriptionConfig(BaseModel):
"""Configuration for audio transcription."""
pass
class Setup(BaseModel):
"""Setup configuration for the Gemini Live session.
Parameters:
model: Model identifier to use.
system_instruction: System instruction for the model. Defaults to None.
tools: List of available tools/functions. Defaults to None.
generation_config: Generation configuration parameters. Defaults to None.
input_audio_transcription: Input audio transcription config. Defaults to None.
output_audio_transcription: Output audio transcription config. Defaults to None.
realtime_input_config: Realtime input configuration. Defaults to None.
"""
model: str
system_instruction: Optional[SystemInstruction] = None
tools: Optional[List[dict]] = None
@@ -137,6 +239,12 @@ class Setup(BaseModel):
class Config(BaseModel):
"""Configuration message for session setup.
Parameters:
setup: Setup configuration for the session.
"""
setup: Setup
@@ -189,36 +297,86 @@ class GroundingMetadata(BaseModel):
class SetupComplete(BaseModel):
"""Indicates that session setup is complete."""
pass
class InlineData(BaseModel):
"""Inline data embedded in server responses.
Parameters:
mimeType: MIME type of the data.
data: Base64-encoded data content.
"""
mimeType: str
data: str
class Part(BaseModel):
"""Part of a server response containing data or text.
Parameters:
inlineData: Inline binary data. Defaults to None.
text: Text content. Defaults to None.
"""
inlineData: Optional[InlineData] = None
text: Optional[str] = None
class ModelTurn(BaseModel):
"""Represents a turn from the model in the conversation.
Parameters:
parts: List of content parts in the model's response.
"""
parts: List[Part]
class ServerContentInterrupted(BaseModel):
"""Indicates server content was interrupted.
Parameters:
interrupted: Whether the content was interrupted.
"""
interrupted: bool
class ServerContentTurnComplete(BaseModel):
"""Indicates the server's turn is complete.
Parameters:
turnComplete: Whether the turn is complete.
"""
turnComplete: bool
class BidiGenerateContentTranscription(BaseModel):
"""Transcription data from bidirectional content generation.
Parameters:
text: The transcribed text content.
"""
text: str
class ServerContent(BaseModel):
"""Content sent from server to client.
Parameters:
modelTurn: Model's conversational turn. Defaults to None.
interrupted: Whether content was interrupted. Defaults to None.
turnComplete: Whether the turn is complete. Defaults to None.
inputTranscription: Transcription of input audio. Defaults to None.
outputTranscription: Transcription of output audio. Defaults to None.
"""
modelTurn: Optional[ModelTurn] = None
interrupted: Optional[bool] = None
turnComplete: Optional[bool] = None
@@ -228,12 +386,26 @@ class ServerContent(BaseModel):
class FunctionCall(BaseModel):
"""Represents a function call from the model.
Parameters:
id: Unique identifier for the function call.
name: Name of the function to call.
args: Arguments to pass to the function.
"""
id: str
name: str
args: dict
class ToolCall(BaseModel):
"""Contains one or more function calls.
Parameters:
functionCalls: List of function calls to execute.
"""
functionCalls: List[FunctionCall]
@@ -248,14 +420,32 @@ class Modality(str, Enum):
class ModalityTokenCount(BaseModel):
"""Token count for a specific modality."""
"""Token count for a specific modality.
Parameters:
modality: The modality type.
tokenCount: Number of tokens for this modality.
"""
modality: Modality
tokenCount: int
class UsageMetadata(BaseModel):
"""Usage metadata about the response."""
"""Usage metadata about the API response.
Parameters:
promptTokenCount: Number of tokens in the prompt. Defaults to None.
cachedContentTokenCount: Number of cached content tokens. Defaults to None.
responseTokenCount: Number of tokens in the response. Defaults to None.
toolUsePromptTokenCount: Number of tokens for tool use prompts. Defaults to None.
thoughtsTokenCount: Number of tokens for model thoughts. Defaults to None.
totalTokenCount: Total number of tokens used. Defaults to None.
promptTokensDetails: Detailed breakdown of prompt tokens by modality. Defaults to None.
cacheTokensDetails: Detailed breakdown of cache tokens by modality. Defaults to None.
responseTokensDetails: Detailed breakdown of response tokens by modality. Defaults to None.
toolUsePromptTokensDetails: Detailed breakdown of tool use tokens by modality. Defaults to None.
"""
promptTokenCount: Optional[int] = None
cachedContentTokenCount: Optional[int] = None
@@ -270,15 +460,32 @@ class UsageMetadata(BaseModel):
class ServerEvent(BaseModel):
"""Server event received from the Gemini Live API.
Parameters:
setupComplete: Setup completion notification. Defaults to None.
serverContent: Content from the server. Defaults to None.
toolCall: Tool/function call request. Defaults to None.
usageMetadata: Token usage metadata. Defaults to None.
"""
setupComplete: Optional[SetupComplete] = None
serverContent: Optional[ServerContent] = None
toolCall: Optional[ToolCall] = None
usageMetadata: Optional[UsageMetadata] = None
def parse_server_event(message_str):
from loguru import logger # Import logger locally to avoid scoping issues
def parse_server_event(str):
"""Parse a server event from JSON string.
Args:
str: JSON string containing the server event.
Returns:
ServerEvent instance if parsing succeeds, None otherwise.
"""
try:
evt_dict = json.loads(message_str)
@@ -301,7 +508,12 @@ def parse_server_event(message_str):
class ContextWindowCompressionConfig(BaseModel):
"""Configuration for context window compression."""
"""Configuration for context window compression.
Parameters:
sliding_window: Whether to use sliding window compression. Defaults to True.
trigger_tokens: Token count threshold to trigger compression. Defaults to None.
"""
sliding_window: Optional[bool] = Field(default=True)
trigger_tokens: Optional[int] = Field(default=None)

247
src/pipecat/services/gemini_multimodal_live/gemini.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google Gemini Multimodal Live API service implementation.
This module provides real-time conversational AI capabilities using Google's
Gemini Multimodal Live API, supporting both text and audio modalities with
voice transcription, streaming responses, and tool usage.
"""
import base64
import json
import time
@@ -62,9 +69,10 @@ from pipecat.services.openai.llm import (
OpenAIUserContextAggregator,
)
from pipecat.transcriptions.language import Language
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.string import match_endofsentence
from pipecat.utils.time import time_now_iso8601
from pipecat.utils.tracing.service_decorators import traced_gemini_live, traced_stt, traced_tts
from pipecat.utils.tracing.service_decorators import traced_gemini_live, traced_stt
from . import events
@@ -88,7 +96,11 @@ def language_to_gemini_language(language: Language) -> Optional[str]:
Source:
https://ai.google.dev/api/generate-content#MediaResolution
Returns None if the language is not supported by Gemini Live.
Args:
language: The language enum value to convert.
Returns:
The Gemini language code string, or None if the language is not supported.
"""
language_map = {
# Arabic
@@ -175,8 +187,22 @@ def language_to_gemini_language(language: Language) -> Optional[str]:
class GeminiMultimodalLiveContext(OpenAILLMContext):
"""Extended OpenAI context for Gemini Multimodal Live API.
Provides Gemini-specific context management including system instruction
extraction and message format conversion for the Live API.
"""
@staticmethod
def upgrade(obj: OpenAILLMContext) -> "GeminiMultimodalLiveContext":
"""Upgrade an OpenAI context to Gemini context.
Args:
obj: The OpenAI context to upgrade.
Returns:
The upgraded Gemini context instance.
"""
if isinstance(obj, OpenAILLMContext) and not isinstance(obj, GeminiMultimodalLiveContext):
logger.debug(f"Upgrading to Gemini Multimodal Live Context: {obj}")
obj.__class__ = GeminiMultimodalLiveContext
@@ -187,6 +213,11 @@ class GeminiMultimodalLiveContext(OpenAILLMContext):
pass
def extract_system_instructions(self):
"""Extract system instructions from context messages.
Returns:
Combined system instruction text from all system messages.
"""
system_instruction = ""
for item in self.messages:
if item.get("role") == "system":
@@ -221,6 +252,11 @@ class GeminiMultimodalLiveContext(OpenAILLMContext):
logger.info(f"Added file reference to context: {file_uri}")
def get_messages_for_initializing_history(self):
"""Get messages formatted for Gemini history initialization.
Returns:
List of messages in Gemini format for conversation history.
"""
messages = []
for item in self.messages:
role = item.get("role")
@@ -256,7 +292,19 @@ class GeminiMultimodalLiveContext(OpenAILLMContext):
class GeminiMultimodalLiveUserContextAggregator(OpenAIUserContextAggregator):
"""User context aggregator for Gemini Multimodal Live.
Extends OpenAI user aggregator to handle Gemini-specific message passing
while maintaining compatibility with the standard aggregation pipeline.
"""
async def process_frame(self, frame, direction):
"""Process incoming frames for user context aggregation.
Args:
frame: The frame to process.
direction: The frame processing direction.
"""
await super().process_frame(frame, direction)
# kind of a hack just to pass the LLMMessagesAppendFrame through, but it's fine for now
if isinstance(frame, LLMMessagesAppendFrame):
@@ -264,15 +312,33 @@ class GeminiMultimodalLiveUserContextAggregator(OpenAIUserContextAggregator):
class GeminiMultimodalLiveAssistantContextAggregator(OpenAIAssistantContextAggregator):
# The LLMAssistantContextAggregator uses TextFrames to aggregate the LLM output,
# but the GeminiMultimodalLiveAssistantContextAggregator pushes LLMTextFrames and TTSTextFrames. We
# need to override this proces_frame for LLMTextFrame, so that only the TTSTextFrames
# are process. This ensures that the context gets only one set of messages.
"""Assistant context aggregator for Gemini Multimodal Live.
Handles assistant response aggregation while filtering out LLMTextFrames
to prevent duplicate context entries, as Gemini Live pushes both
LLMTextFrames and TTSTextFrames.
"""
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames for assistant context aggregation.
Args:
frame: The frame to process.
direction: The frame processing direction.
"""
# The LLMAssistantContextAggregator uses TextFrames to aggregate the LLM output,
# but the GeminiMultimodalLiveAssistantContextAggregator pushes LLMTextFrames and TTSTextFrames. We
# need to override this proces_frame for LLMTextFrame, so that only the TTSTextFrames
# are process. This ensures that the context gets only one set of messages.
if not isinstance(frame, LLMTextFrame):
await super().process_frame(frame, direction)
async def handle_user_image_frame(self, frame: UserImageRawFrame):
"""Handle user image frames.
Args:
frame: The user image frame to handle.
"""
# We don't want to store any images in the context. Revisit this later
# when the API evolves.
pass
@@ -280,17 +346,41 @@ class GeminiMultimodalLiveAssistantContextAggregator(OpenAIAssistantContextAggre
@dataclass
class GeminiMultimodalLiveContextAggregatorPair:
"""Pair of user and assistant context aggregators for Gemini Multimodal Live.
Parameters:
_user: The user context aggregator instance.
_assistant: The assistant context aggregator instance.
"""
_user: GeminiMultimodalLiveUserContextAggregator
_assistant: GeminiMultimodalLiveAssistantContextAggregator
def user(self) -> GeminiMultimodalLiveUserContextAggregator:
"""Get the user context aggregator.
Returns:
The user context aggregator instance.
"""
return self._user
def assistant(self) -> GeminiMultimodalLiveAssistantContextAggregator:
"""Get the assistant context aggregator.
Returns:
The assistant context aggregator instance.
"""
return self._assistant
class GeminiMultimodalModalities(Enum):
"""Supported modalities for Gemini Multimodal Live.
Parameters:
TEXT: Text responses.
AUDIO: Audio responses.
"""
TEXT = "TEXT"
AUDIO = "AUDIO"
@@ -305,7 +395,15 @@ class GeminiMediaResolution(str, Enum):
class GeminiVADParams(BaseModel):
"""Voice Activity Detection parameters."""
"""Voice Activity Detection parameters for Gemini Live.
Parameters:
disabled: Whether to disable VAD. Defaults to None.
start_sensitivity: Sensitivity for speech start detection. Defaults to None.
end_sensitivity: Sensitivity for speech end detection. Defaults to None.
prefix_padding_ms: Prefix padding in milliseconds. Defaults to None.
silence_duration_ms: Silence duration threshold in milliseconds. Defaults to None.
"""
disabled: Optional[bool] = Field(default=None)
start_sensitivity: Optional[events.StartSensitivity] = Field(default=None)
@@ -315,7 +413,12 @@ class GeminiVADParams(BaseModel):
class ContextWindowCompressionParams(BaseModel):
"""Parameters for context window compression."""
"""Parameters for context window compression in Gemini Live.
Parameters:
enabled: Whether compression is enabled. Defaults to False.
trigger_tokens: Token count to trigger compression. None uses 80% of context window.
"""
enabled: bool = Field(default=False)
trigger_tokens: Optional[int] = Field(
@@ -324,6 +427,23 @@ class ContextWindowCompressionParams(BaseModel):
class InputParams(BaseModel):
"""Input parameters for Gemini Multimodal Live generation.
Parameters:
frequency_penalty: Frequency penalty for generation (0.0-2.0). Defaults to None.
max_tokens: Maximum tokens to generate. Must be >= 1. Defaults to 4096.
presence_penalty: Presence penalty for generation (0.0-2.0). Defaults to None.
temperature: Sampling temperature (0.0-2.0). Defaults to None.
top_k: Top-k sampling parameter. Must be >= 0. Defaults to None.
top_p: Top-p sampling parameter (0.0-1.0). Defaults to None.
modalities: Response modalities. Defaults to AUDIO.
language: Language for generation. Defaults to EN_US.
media_resolution: Media resolution setting. Defaults to UNSPECIFIED.
vad: Voice activity detection parameters. Defaults to None.
context_window_compression: Context compression settings. Defaults to None.
extra: Additional parameters. Defaults to empty dict.
"""
frequency_penalty: Optional[float] = Field(default=None, ge=0.0, le=2.0)
max_tokens: Optional[int] = Field(default=4096, ge=1)
presence_penalty: Optional[float] = Field(default=None, ge=0.0, le=2.0)
@@ -348,25 +468,6 @@ class GeminiMultimodalLiveLLMService(LLMService):
This service enables real-time conversations with Gemini, supporting both
text and audio modalities. It handles voice transcription, streaming audio
responses, and tool usage.
Args:
api_key (str): Google AI API key
base_url (str, optional): API endpoint base URL. Defaults to
"generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent".
model (str, optional): Model identifier to use. Defaults to
"models/gemini-2.0-flash-live-001".
voice_id (str, optional): TTS voice identifier. Defaults to "Charon".
start_audio_paused (bool, optional): Whether to start with audio input paused.
Defaults to False.
start_video_paused (bool, optional): Whether to start with video input paused.
Defaults to False.
system_instruction (str, optional): System prompt for the model. Defaults to None.
tools (Union[List[dict], ToolsSchema], optional): Tools/functions available to the model.
Defaults to None.
params (InputParams, optional): Configuration parameters for the model.
Defaults to InputParams().
inference_on_context_initialization (bool, optional): Whether to generate a response
when context is first set. Defaults to True.
"""
# Overriding the default adapter to use the Gemini one.
@@ -388,6 +489,22 @@ class GeminiMultimodalLiveLLMService(LLMService):
file_api_base_url: str = "https://generativelanguage.googleapis.com/v1beta/files",
**kwargs,
):
"""Initialize the Gemini Multimodal Live LLM service.
Args:
api_key: Google AI API key for authentication.
base_url: API endpoint base URL. Defaults to the official Gemini Live endpoint.
model: Model identifier to use. Defaults to "models/gemini-2.0-flash-live-001".
voice_id: TTS voice identifier. Defaults to "Charon".
start_audio_paused: Whether to start with audio input paused. Defaults to False.
start_video_paused: Whether to start with video input paused. Defaults to False.
system_instruction: System prompt for the model. Defaults to None.
tools: Tools/functions available to the model. Defaults to None.
params: Configuration parameters for the model. Defaults to InputParams().
inference_on_context_initialization: Whether to generate a response when context
is first set. Defaults to True.
**kwargs: Additional arguments passed to parent LLMService.
"""
super().__init__(base_url=base_url, **kwargs)
params = params or InputParams()
@@ -456,19 +573,43 @@ class GeminiMultimodalLiveLLMService(LLMService):
self._accumulated_grounding_metadata = None
def can_generate_metrics(self) -> bool:
"""Check if the service can generate usage metrics.
Returns:
True as Gemini Live supports token usage metrics.
"""
return True
def set_audio_input_paused(self, paused: bool):
"""Set the audio input pause state.
Args:
paused: Whether to pause audio input.
"""
self._audio_input_paused = paused
def set_video_input_paused(self, paused: bool):
"""Set the video input pause state.
Args:
paused: Whether to pause video input.
"""
self._video_input_paused = paused
def set_model_modalities(self, modalities: GeminiMultimodalModalities):
"""Set the model response modalities.
Args:
modalities: The modalities to use for responses.
"""
self._settings["modalities"] = modalities
def set_language(self, language: Language):
"""Set the language for generation."""
"""Set the language for generation.
Args:
language: The language to use for generation.
"""
self._language = language
self._language_code = language_to_gemini_language(language) or "en-US"
self._settings["language"] = self._language_code
@@ -481,6 +622,9 @@ class GeminiMultimodalLiveLLMService(LLMService):
way to trigger the pipeline. This sends the history to the server. The `inference_on_context_initialization`
flag controls whether to set the turnComplete flag when we do this. Without that flag, the model will
not respond. This is often what we want when setting the context at the beginning of a conversation.
Args:
context: The OpenAI LLM context to set.
"""
if self._context:
logger.error(
@@ -495,14 +639,29 @@ class GeminiMultimodalLiveLLMService(LLMService):
#
async def start(self, frame: StartFrame):
"""Start the service and establish websocket connection.
Args:
frame: The start frame.
"""
await super().start(frame)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the service and close connections.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the service and close connections.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._disconnect()
@@ -537,6 +696,12 @@ class GeminiMultimodalLiveLLMService(LLMService):
#
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames for the Gemini Live service.
Args:
frame: The frame to process.
direction: The frame processing direction.
"""
await super().process_frame(frame, direction)
if isinstance(frame, TranscriptionFrame):
@@ -592,6 +757,11 @@ class GeminiMultimodalLiveLLMService(LLMService):
#
async def send_client_event(self, event):
"""Send a client event to the Gemini Live API.
Args:
event: The event to send.
"""
await self._ws_send(event.model_dump(exclude_none=True))
async def _connect(self):
@@ -735,9 +905,7 @@ class GeminiMultimodalLiveLLMService(LLMService):
#
async def _receive_task_handler(self):
async for message in self._websocket:
self.start_watchdog()
async for message in WatchdogAsyncIterator(self._websocket, manager=self.task_manager):
evt = events.parse_server_event(message)
# logger.debug(f"Received event: {message[:500]}")
# logger.debug(f"Received event: {evt}")
@@ -767,8 +935,6 @@ class GeminiMultimodalLiveLLMService(LLMService):
pass
self.reset_watchdog()
#
#
#
@@ -1186,22 +1352,19 @@ class GeminiMultimodalLiveLLMService(LLMService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> GeminiMultimodalLiveContextAggregatorPair:
"""Create an instance of GeminiMultimodalLiveContextAggregatorPair from
an OpenAILLMContext. Constructor keyword arguments for both the user and
assistant aggregators can be provided.
"""Create an instance of GeminiMultimodalLiveContextAggregatorPair from an OpenAILLMContext.
Constructor keyword arguments for both the user and assistant aggregators can be provided.
Args:
context (OpenAILLMContext): The LLM context.
user_params (LLMUserAggregatorParams, optional): User aggregator
parameters.
assistant_params (LLMAssistantAggregatorParams, optional): User
aggregator parameters.
context: The LLM context to use.
user_params: User aggregator parameters. Defaults to LLMUserAggregatorParams().
assistant_params: Assistant aggregator parameters. Defaults to LLMAssistantAggregatorParams().
Returns:
GeminiMultimodalLiveContextAggregatorPair: A pair of context
aggregators, one for the user and one for the assistant,
encapsulated in an GeminiMultimodalLiveContextAggregatorPair.
"""
context.set_llm_adapter(self.get_llm_adapter())

20
src/pipecat/services/gladia/config.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Configuration for the Gladia STT service."""
from typing import Any, Dict, List, Optional, Union
from pydantic import BaseModel
@@ -14,7 +16,7 @@ from pipecat.transcriptions.language import Language
class LanguageConfig(BaseModel):
"""Configuration for language detection and handling.
Attributes:
Parameters:
languages: List of language codes to use for transcription
code_switching: Whether to auto-detect language changes during transcription
"""
@@ -26,7 +28,7 @@ class LanguageConfig(BaseModel):
class PreProcessingConfig(BaseModel):
"""Configuration for audio pre-processing options.
Attributes:
Parameters:
speech_threshold: Sensitivity for speech detection (0-1)
"""
@@ -36,7 +38,7 @@ class PreProcessingConfig(BaseModel):
class CustomVocabularyItem(BaseModel):
"""Represents a custom vocabulary item with an intensity value.
Attributes:
Parameters:
value: The vocabulary word or phrase
intensity: The bias intensity for this vocabulary item (0-1)
"""
@@ -48,7 +50,7 @@ class CustomVocabularyItem(BaseModel):
class CustomVocabularyConfig(BaseModel):
"""Configuration for custom vocabulary.
Attributes:
Parameters:
vocabulary: List of words/phrases or CustomVocabularyItem objects
default_intensity: Default intensity for simple string vocabulary items
"""
@@ -60,7 +62,7 @@ class CustomVocabularyConfig(BaseModel):
class CustomSpellingConfig(BaseModel):
"""Configuration for custom spelling rules.
Attributes:
Parameters:
spelling_dictionary: Mapping of correct spellings to phonetic variations
"""
@@ -70,7 +72,7 @@ class CustomSpellingConfig(BaseModel):
class TranslationConfig(BaseModel):
"""Configuration for real-time translation.
Attributes:
Parameters:
target_languages: List of target language codes for translation
model: Translation model to use ("base" or "enhanced")
match_original_utterances: Whether to align translations with original utterances
@@ -92,7 +94,7 @@ class TranslationConfig(BaseModel):
class RealtimeProcessingConfig(BaseModel):
"""Configuration for real-time processing features.
Attributes:
Parameters:
words_accurate_timestamps: Whether to provide per-word timestamps
custom_vocabulary: Whether to enable custom vocabulary
custom_vocabulary_config: Custom vocabulary configuration
@@ -118,7 +120,7 @@ class RealtimeProcessingConfig(BaseModel):
class MessagesConfig(BaseModel):
"""Configuration for controlling which message types are sent via WebSocket.
Attributes:
Parameters:
receive_partial_transcripts: Whether to receive intermediate transcription results
receive_final_transcripts: Whether to receive final transcription results
receive_speech_events: Whether to receive speech begin/end events
@@ -144,7 +146,7 @@ class MessagesConfig(BaseModel):
class GladiaInputParams(BaseModel):
"""Configuration parameters for the Gladia STT service.
Attributes:
Parameters:
encoding: Audio encoding format
bit_depth: Audio bit depth
channels: Number of audio channels

93
src/pipecat/services/gladia/stt.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Gladia Speech-to-Text (STT) service implementation.
This module provides a Speech-to-Text service using Gladia's real-time WebSocket API,
supporting multiple languages, custom vocabulary, and various audio processing options.
"""
import asyncio
import base64
import json
@@ -25,6 +31,7 @@ from pipecat.frames.frames import (
from pipecat.services.gladia.config import GladiaInputParams
from pipecat.services.stt_service import STTService
from pipecat.transcriptions.language import Language
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.time import time_now_iso8601
from pipecat.utils.tracing.service_decorators import traced_stt
@@ -40,10 +47,10 @@ def language_to_gladia_language(language: Language) -> Optional[str]:
"""Convert a Language enum to Gladia's language code format.
Args:
language: The Language enum value to convert
language: The Language enum value to convert.
Returns:
The Gladia language code string or None if not supported
The Gladia language code string or None if not supported.
"""
BASE_LANGUAGES = {
Language.AF: "af",
@@ -179,6 +186,7 @@ class GladiaSTTService(STTService):
This service connects to Gladia's WebSocket API for real-time transcription
with support for multiple languages, custom vocabulary, and various processing options.
Provides automatic reconnection, audio buffering, and comprehensive error handling.
For complete API documentation, see: https://docs.gladia.io/api-reference/v2/live/init
"""
@@ -203,16 +211,16 @@ class GladiaSTTService(STTService):
"""Initialize the Gladia STT service.
Args:
api_key: Gladia API key
url: Gladia API URL
confidence: Minimum confidence threshold for transcriptions
sample_rate: Audio sample rate in Hz
model: Model to use ("solaria-1")
params: Additional configuration parameters
max_reconnection_attempts: Maximum number of reconnection attempts
reconnection_delay: Initial delay between reconnection attempts (exponential backoff)
max_buffer_size: Maximum size of audio buffer in bytes
**kwargs: Additional arguments passed to the STTService
api_key: Gladia API key for authentication.
url: Gladia API URL. Defaults to "https://api.gladia.io/v2/live".
confidence: Minimum confidence threshold for transcriptions (0.0-1.0).
sample_rate: Audio sample rate in Hz. If None, uses service default.
model: Model to use for transcription. Defaults to "solaria-1".
params: Additional configuration parameters for Gladia service.
max_reconnection_attempts: Maximum number of reconnection attempts. Defaults to 5.
reconnection_delay: Initial delay between reconnection attempts in seconds.
max_buffer_size: Maximum size of audio buffer in bytes. Defaults to 20MB.
**kwargs: Additional arguments passed to the STTService parent class.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
@@ -255,10 +263,22 @@ class GladiaSTTService(STTService):
self._should_reconnect = True
def can_generate_metrics(self) -> bool:
"""Check if the service can generate performance metrics.
Returns:
True, indicating this service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert pipecat Language enum to Gladia's language code."""
"""Convert pipecat Language enum to Gladia's language code.
Args:
language: The Language enum value to convert.
Returns:
The Gladia language code string or None if not supported.
"""
return language_to_gladia_language(language)
def _prepare_settings(self) -> Dict[str, Any]:
@@ -313,7 +333,11 @@ class GladiaSTTService(STTService):
return settings
async def start(self, frame: StartFrame):
"""Start the Gladia STT websocket connection."""
"""Start the Gladia STT websocket connection.
Args:
frame: The start frame triggering service startup.
"""
await super().start(frame)
if self._connection_task:
return
@@ -322,7 +346,11 @@ class GladiaSTTService(STTService):
self._connection_task = self.create_task(self._connection_handler())
async def stop(self, frame: EndFrame):
"""Stop the Gladia STT websocket connection."""
"""Stop the Gladia STT websocket connection.
Args:
frame: The end frame triggering service shutdown.
"""
await super().stop(frame)
self._should_reconnect = False
await self._send_stop_recording()
@@ -334,7 +362,11 @@ class GladiaSTTService(STTService):
await self._cleanup_connection()
async def cancel(self, frame: CancelFrame):
"""Cancel the Gladia STT websocket connection."""
"""Cancel the Gladia STT websocket connection.
Args:
frame: The cancel frame triggering service cancellation.
"""
await super().cancel(frame)
self._should_reconnect = False
@@ -345,7 +377,14 @@ class GladiaSTTService(STTService):
await self._cleanup_connection()
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Run speech-to-text on audio data."""
"""Run speech-to-text on audio data.
Args:
audio: Raw audio bytes to transcribe.
Yields:
None (processing is handled asynchronously via WebSocket).
"""
await self.start_ttfb_metrics()
await self.start_processing_metrics()
@@ -391,8 +430,8 @@ class GladiaSTTService(STTService):
await self._send_buffered_audio()
# Start tasks
self._receive_task = asyncio.create_task(self._receive_task_handler())
self._keepalive_task = asyncio.create_task(self._keepalive_task_handler())
self._receive_task = self.create_task(self._receive_task_handler())
self._keepalive_task = self.create_task(self._keepalive_task_handler())
# Wait for tasks to complete
await asyncio.gather(self._receive_task, self._keepalive_task)
@@ -403,9 +442,9 @@ class GladiaSTTService(STTService):
# Clean up tasks
if self._receive_task:
self._receive_task.cancel()
await self.cancel_task(self._receive_task)
if self._keepalive_task:
self._keepalive_task.cancel()
await self.cancel_task(self._keepalive_task)
# Attempt reconnect using helper
if not await self._maybe_reconnect():
@@ -484,9 +523,11 @@ class GladiaSTTService(STTService):
async def _keepalive_task_handler(self):
"""Send periodic empty audio chunks to keep the connection alive."""
try:
KEEPALIVE_SLEEP = 20 if self.task_manager.task_watchdog_enabled else 3
while self._connection_active:
# Send keepalive every 20 seconds (Gladia times out after 30 seconds)
await asyncio.sleep(20)
self.reset_watchdog()
# Send keepalive (Gladia times out after 30 seconds)
await asyncio.sleep(KEEPALIVE_SLEEP)
if self._websocket and not self._websocket.closed:
# Send an empty audio chunk as keepalive
empty_audio = b""
@@ -501,9 +542,7 @@ class GladiaSTTService(STTService):
async def _receive_task_handler(self):
try:
async for message in self._websocket:
self.start_watchdog()
async for message in WatchdogAsyncIterator(self._websocket, manager=self.task_manager):
content = json.loads(message)
# Handle audio chunk acknowledgments
@@ -568,8 +607,6 @@ class GladiaSTTService(STTService):
pass
except Exception as e:
logger.error(f"Error in Gladia WebSocket handler: {e}")
finally:
self.reset_watchdog()
async def _maybe_reconnect(self) -> bool:
"""Handle exponential backoff reconnection logic."""

40
src/pipecat/services/google/frames.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google AI service frames for search and grounding functionality.
This module defines specialized frame types for handling search results
and grounding metadata from Google AI models, particularly for Gemini
models that support web search and fact grounding capabilities.
"""
from dataclasses import dataclass, field
from typing import List, Optional
@@ -12,12 +19,27 @@ from pipecat.frames.frames import DataFrame
@dataclass
class LLMSearchResult:
"""Represents a single search result with confidence scores.
Parameters:
text: The search result text content.
confidence: List of confidence scores associated with the result.
"""
text: str
confidence: List[float] = field(default_factory=list)
@dataclass
class LLMSearchOrigin:
"""Represents the origin source of search results.
Parameters:
site_uri: URI of the source website.
site_title: Title of the source website.
results: List of search results from this origin.
"""
site_uri: Optional[str] = None
site_title: Optional[str] = None
results: List[LLMSearchResult] = field(default_factory=list)
@@ -25,9 +47,27 @@ class LLMSearchOrigin:
@dataclass
class LLMSearchResponseFrame(DataFrame):
"""Frame containing search results and grounding information from Google AI models.
This frame is used to convey search results and grounding metadata
from Google AI models that support web search capabilities. It includes
the search result text, rendered content, and detailed origin information
with confidence scores.
Parameters:
search_result: The main search result text.
rendered_content: Rendered content from the search entry point.
origins: List of search result origins with detailed information.
"""
search_result: Optional[str] = None
rendered_content: Optional[str] = None
origins: List[LLMSearchOrigin] = field(default_factory=list)
def __str__(self):
"""Return string representation of the search response frame.
Returns:
String representation showing search result and origins.
"""
return f"LLMSearchResponseFrame(search_result={self.search_result}, origins={self.origins})"

41
src/pipecat/services/google/image.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google AI image generation service implementation.
This module provides integration with Google's Imagen model for generating
images from text prompts using the Google AI API.
"""
import io
import os
@@ -29,7 +35,22 @@ except ModuleNotFoundError as e:
class GoogleImageGenService(ImageGenService):
"""Google AI image generation service using Imagen models.
Provides text-to-image generation capabilities using Google's Imagen models
through the Google AI API. Supports multiple image generation and negative
prompting for enhanced control over generated content.
"""
class InputParams(BaseModel):
"""Configuration parameters for Google image generation.
Parameters:
number_of_images: Number of images to generate (1-8). Defaults to 1.
model: Google Imagen model to use. Defaults to "imagen-3.0-generate-002".
negative_prompt: Optional negative prompt to guide what not to include.
"""
number_of_images: int = Field(default=1, ge=1, le=8)
model: str = Field(default="imagen-3.0-generate-002")
negative_prompt: Optional[str] = Field(default=None)
@@ -41,22 +62,38 @@ class GoogleImageGenService(ImageGenService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the GoogleImageGenService with API key and parameters.
Args:
api_key: Google AI API key for authentication.
params: Configuration parameters for image generation. Defaults to InputParams().
**kwargs: Additional arguments passed to the parent ImageGenService.
"""
super().__init__(**kwargs)
self._params = params or GoogleImageGenService.InputParams()
self._client = genai.Client(api_key=api_key)
self.set_model_name(self._params.model)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Google image generation service supports metrics.
"""
return True
async def run_image_gen(self, prompt: str) -> AsyncGenerator[Frame, None]:
"""Generate images from a text prompt using Google's Imagen model.
Args:
prompt (str): The text description to generate images from.
prompt: The text description to generate images from.
Yields:
Frame: Generated image frames or error frames.
Frame: Generated URLImageRawFrame objects containing the generated
images, or ErrorFrame objects if generation fails.
Raises:
Exception: If there are issues with the Google AI API or image processing.
"""
logger.debug(f"Generating image from prompt: {prompt}")
await self.start_ttfb_metrics()

196
src/pipecat/services/google/llm.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google Gemini integration for Pipecat.
This module provides Google Gemini integration for the Pipecat framework,
including LLM services, context management, and message aggregation.
"""
import base64
import io
import json
@@ -47,6 +53,7 @@ from pipecat.services.openai.llm import (
OpenAIAssistantContextAggregator,
OpenAIUserContextAggregator,
)
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.tracing.service_decorators import traced_llm
# Suppress gRPC fork warnings
@@ -70,7 +77,14 @@ except ModuleNotFoundError as e:
class GoogleUserContextAggregator(OpenAIUserContextAggregator):
"""Google-specific user context aggregator.
Extends OpenAI user context aggregator to handle Google AI's specific
Content and Part message format for user messages.
"""
async def push_aggregation(self):
"""Push aggregated user text as a Google Content message."""
if len(self._aggregation) > 0:
self._context.add_message(Content(role="user", parts=[Part(text=self._aggregation)]))
@@ -87,10 +101,26 @@ class GoogleUserContextAggregator(OpenAIUserContextAggregator):
class GoogleAssistantContextAggregator(OpenAIAssistantContextAggregator):
"""Google-specific assistant context aggregator.
Extends OpenAI assistant context aggregator to handle Google AI's specific
Content and Part message format for assistant responses and function calls.
"""
async def handle_aggregation(self, aggregation: str):
"""Handle aggregated assistant text response.
Args:
aggregation: The aggregated text response from the assistant.
"""
self._context.add_message(Content(role="model", parts=[Part(text=aggregation)]))
async def handle_function_call_in_progress(self, frame: FunctionCallInProgressFrame):
"""Handle function call in progress frame.
Args:
frame: Frame containing function call details.
"""
self._context.add_message(
Content(
role="model",
@@ -119,6 +149,11 @@ class GoogleAssistantContextAggregator(OpenAIAssistantContextAggregator):
)
async def handle_function_call_result(self, frame: FunctionCallResultFrame):
"""Handle function call result frame.
Args:
frame: Frame containing function call result.
"""
if frame.result:
await self._update_function_call_result(
frame.function_name, frame.tool_call_id, frame.result
@@ -129,6 +164,11 @@ class GoogleAssistantContextAggregator(OpenAIAssistantContextAggregator):
)
async def handle_function_call_cancel(self, frame: FunctionCallCancelFrame):
"""Handle function call cancellation frame.
Args:
frame: Frame containing function call cancellation details.
"""
await self._update_function_call_result(
frame.function_name, frame.tool_call_id, "CANCELLED"
)
@@ -143,6 +183,11 @@ class GoogleAssistantContextAggregator(OpenAIAssistantContextAggregator):
part.function_response.response = {"value": json.dumps(result)}
async def handle_user_image_frame(self, frame: UserImageRawFrame):
"""Handle user image frame.
Args:
frame: Frame containing user image data and request context.
"""
await self._update_function_call_result(
frame.request.function_name, frame.request.tool_call_id, "COMPLETED"
)
@@ -156,28 +201,66 @@ class GoogleAssistantContextAggregator(OpenAIAssistantContextAggregator):
@dataclass
class GoogleContextAggregatorPair:
"""Pair of Google context aggregators for user and assistant messages.
Parameters:
_user: User context aggregator for handling user messages.
_assistant: Assistant context aggregator for handling assistant responses.
"""
_user: GoogleUserContextAggregator
_assistant: GoogleAssistantContextAggregator
def user(self) -> GoogleUserContextAggregator:
"""Get the user context aggregator.
Returns:
The user context aggregator instance.
"""
return self._user
def assistant(self) -> GoogleAssistantContextAggregator:
"""Get the assistant context aggregator.
Returns:
The assistant context aggregator instance.
"""
return self._assistant
class GoogleLLMContext(OpenAILLMContext):
"""Google AI LLM context that extends OpenAI context for Google-specific formatting.
This class handles conversion between OpenAI-style messages and Google AI's
Content/Part format, including system messages, function calls, and media.
"""
def __init__(
self,
messages: Optional[List[dict]] = None,
tools: Optional[List[dict]] = None,
tool_choice: Optional[dict] = None,
):
"""Initialize GoogleLLMContext.
Args:
messages: Initial messages in OpenAI format.
tools: Available tools/functions for the model.
tool_choice: Tool choice configuration.
"""
super().__init__(messages=messages, tools=tools, tool_choice=tool_choice)
self.system_message = None
@staticmethod
def upgrade_to_google(obj: OpenAILLMContext) -> "GoogleLLMContext":
"""Upgrade an OpenAI context to a Google context.
Args:
obj: OpenAI LLM context to upgrade.
Returns:
GoogleLLMContext instance with converted messages.
"""
if isinstance(obj, OpenAILLMContext) and not isinstance(obj, GoogleLLMContext):
logger.debug(f"Upgrading to Google: {obj}")
obj.__class__ = GoogleLLMContext
@@ -185,10 +268,20 @@ class GoogleLLMContext(OpenAILLMContext):
return obj
def set_messages(self, messages: List):
"""Set messages and restructure them for Google format.
Args:
messages: List of messages to set.
"""
self._messages[:] = messages
self._restructure_from_openai_messages()
def add_messages(self, messages: List):
"""Add messages to the context, converting to Google format as needed.
Args:
messages: List of messages to add (can be mixed formats).
"""
# Convert each message individually
converted_messages = []
for msg in messages:
@@ -205,6 +298,11 @@ class GoogleLLMContext(OpenAILLMContext):
self._messages.extend(converted_messages)
def get_messages_for_logging(self):
"""Get messages formatted for logging with sensitive data redacted.
Returns:
List of message dictionaries with inline data redacted.
"""
msgs = []
for message in self.messages:
obj = message.to_json_dict()
@@ -221,6 +319,14 @@ class GoogleLLMContext(OpenAILLMContext):
def add_image_frame_message(
self, *, format: str, size: tuple[int, int], image: bytes, text: str = None
):
"""Add an image message to the context.
Args:
format: Image format (e.g., 'RGB', 'RGBA').
size: Image dimensions as (width, height).
image: Raw image bytes.
text: Optional text to accompany the image.
"""
buffer = io.BytesIO()
Image.frombytes(format, size, image).save(buffer, format="JPEG")
@@ -234,6 +340,12 @@ class GoogleLLMContext(OpenAILLMContext):
def add_audio_frames_message(
self, *, audio_frames: list[AudioRawFrame], text: str = "Audio follows"
):
"""Add audio frames as a message to the context.
Args:
audio_frames: List of audio frames to add.
text: Text description of the audio content.
"""
if not audio_frames:
return
@@ -447,17 +559,28 @@ class GoogleLLMContext(OpenAILLMContext):
class GoogleLLMService(LLMService):
"""This class implements inference with Google's AI models.
"""Google AI (Gemini) LLM service implementation.
This service translates internally from OpenAILLMContext to the messages format
expected by the Google AI model. We are using the OpenAILLMContext as a lingua
franca for all LLM services, so that it is easy to switch between different LLMs.
This class implements inference with Google's AI models, translating internally
from OpenAILLMContext to the messages format expected by the Google AI model.
We use OpenAILLMContext as a lingua franca for all LLM services to enable
easy switching between different LLMs.
"""
# Overriding the default adapter to use the Gemini one.
adapter_class = GeminiLLMAdapter
class InputParams(BaseModel):
"""Input parameters for Google AI models.
Parameters:
max_tokens: Maximum number of tokens to generate.
temperature: Sampling temperature between 0.0 and 2.0.
top_k: Top-k sampling parameter.
top_p: Top-p sampling parameter between 0.0 and 1.0.
extra: Additional parameters as a dictionary.
"""
max_tokens: Optional[int] = Field(default=4096, ge=1)
temperature: Optional[float] = Field(default=None, ge=0.0, le=2.0)
top_k: Optional[int] = Field(default=None, ge=0)
@@ -475,6 +598,17 @@ class GoogleLLMService(LLMService):
tool_config: Optional[Dict[str, Any]] = None,
**kwargs,
):
"""Initialize the Google LLM service.
Args:
api_key: Google AI API key for authentication.
model: Model name to use. Defaults to "gemini-2.0-flash".
params: Input parameters for the model.
system_instruction: System instruction/prompt for the model.
tools: List of available tools/functions.
tool_config: Configuration for tool usage.
**kwargs: Additional arguments passed to parent class.
"""
super().__init__(**kwargs)
params = params or GoogleLLMService.InputParams()
@@ -494,11 +628,30 @@ class GoogleLLMService(LLMService):
self._tool_config = tool_config
def can_generate_metrics(self) -> bool:
"""Check if the service can generate usage metrics.
Returns:
True, as Google AI provides token usage metrics.
"""
return True
def _create_client(self, api_key: str):
self._client = genai.Client(api_key=api_key)
def _maybe_unset_thinking_budget(self, generation_params: Dict[str, Any]):
try:
# There's no way to introspect on model capabilities, so
# to check for models that we know default to thinkin on
# and can be configured to turn it off.
if not self._model_name.startswith("gemini-2.5-flash"):
return
# If thinking_config is already set, don't override it.
if "thinking_config" in generation_params:
return
generation_params.setdefault("thinking_config", {})["thinking_budget"] = 0
except Exception as e:
logger.exception(f"Failed to unset thinking budget: {e}")
@traced_llm
async def _process_context(self, context: OpenAILLMContext):
await self.push_frame(LLMFullResponseStartFrame())
@@ -506,6 +659,8 @@ class GoogleLLMService(LLMService):
prompt_tokens = 0
completion_tokens = 0
total_tokens = 0
cache_read_input_tokens = 0
reasoning_tokens = 0
grounding_metadata = None
search_result = ""
@@ -545,6 +700,12 @@ class GoogleLLMService(LLMService):
if v is not None
}
if self._settings["extra"]:
generation_params.update(self._settings["extra"])
# possibly modify generation_params (in place) to set thinking to off by default
self._maybe_unset_thinking_budget(generation_params)
generation_config = (
GenerateContentConfig(**generation_params) if generation_params else None
)
@@ -557,13 +718,15 @@ class GoogleLLMService(LLMService):
)
function_calls = []
async for chunk in response:
async for chunk in WatchdogAsyncIterator(response, manager=self.task_manager):
# Stop TTFB metrics after the first chunk
await self.stop_ttfb_metrics()
if chunk.usage_metadata:
prompt_tokens += chunk.usage_metadata.prompt_token_count or 0
completion_tokens += chunk.usage_metadata.candidates_token_count or 0
total_tokens += chunk.usage_metadata.total_token_count or 0
cache_read_input_tokens += chunk.usage_metadata.cached_content_token_count or 0
reasoning_tokens += chunk.usage_metadata.thoughts_token_count or 0
if not chunk.candidates:
continue
@@ -645,11 +808,19 @@ class GoogleLLMService(LLMService):
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=total_tokens,
cache_read_input_tokens=cache_read_input_tokens,
reasoning_tokens=reasoning_tokens,
)
)
await self.push_frame(LLMFullResponseEndFrame())
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process incoming frames and handle different frame types.
Args:
frame: The frame to process.
direction: Direction of frame processing.
"""
await super().process_frame(frame, direction)
context = None
@@ -678,16 +849,15 @@ class GoogleLLMService(LLMService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> GoogleContextAggregatorPair:
"""Create an instance of GoogleContextAggregatorPair from an
OpenAILLMContext. Constructor keyword arguments for both the user and
assistant aggregators can be provided.
"""Create Google-specific context aggregators.
Creates a pair of context aggregators optimized for Google's message format,
including support for function calls, tool usage, and image handling.
Args:
context (OpenAILLMContext): The LLM context.
user_params (LLMUserAggregatorParams, optional): User aggregator
parameters.
assistant_params (LLMAssistantAggregatorParams, optional): User
aggregator parameters.
context: The LLM context to create aggregators for.
user_params: Parameters for user message aggregation.
assistant_params: Parameters for assistant message aggregation.
Returns:
GoogleContextAggregatorPair: A pair of context aggregators, one for

30
src/pipecat/services/google/llm_openai.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google LLM service using OpenAI-compatible API format.
This module provides integration with Google's AI LLM models using the OpenAI
API format through Google's Gemini API OpenAI compatibility layer.
"""
import json
import os
@@ -11,6 +17,7 @@ from openai import AsyncStream
from openai.types.chat import ChatCompletionChunk
from pipecat.services.llm_service import FunctionCallFromLLM
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
# Suppress gRPC fork warnings
os.environ["GRPC_ENABLE_FORK_SUPPORT"] = "false"
@@ -24,8 +31,17 @@ from pipecat.services.openai.llm import OpenAILLMService
class GoogleLLMOpenAIBetaService(OpenAILLMService):
"""This class implements inference with Google's AI LLM models using the OpenAI format.
Ref - https://ai.google.dev/gemini-api/docs/openai
"""Google LLM service using OpenAI-compatible API format.
This service provides access to Google's AI LLM models (like Gemini) through
the OpenAI API format. It handles streaming responses, function calls, and
tool usage while maintaining compatibility with OpenAI's interface.
Note: This service includes a workaround for a Google API bug where function
call indices may be incorrectly set to None, resulting in empty function names.
Reference:
https://ai.google.dev/gemini-api/docs/openai
"""
def __init__(
@@ -36,6 +52,14 @@ class GoogleLLMOpenAIBetaService(OpenAILLMService):
model: str = "gemini-2.0-flash",
**kwargs,
):
"""Initialize the Google LLM service.
Args:
api_key: Google API key for authentication.
base_url: Base URL for Google's OpenAI-compatible API.
model: Google model name to use (e.g., "gemini-2.0-flash").
**kwargs: Additional arguments passed to the parent OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
async def _process_context(self, context: OpenAILLMContext):
@@ -53,7 +77,7 @@ class GoogleLLMOpenAIBetaService(OpenAILLMService):
context
)
async for chunk in chunk_stream:
async for chunk in WatchdogAsyncIterator(chunk_stream, manager=self.task_manager):
if chunk.usage:
tokens = LLMTokenUsage(
prompt_tokens=chunk.usage.prompt_tokens,

52
src/pipecat/services/google/llm_vertex.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google Vertex AI LLM service implementation.
This module provides integration with Google's AI models via Vertex AI while
maintaining OpenAI API compatibility through Google's OpenAI-compatible endpoint.
"""
import json
import os
@@ -31,16 +37,24 @@ except ModuleNotFoundError as e:
class GoogleVertexLLMService(OpenAILLMService):
"""Implements inference with Google's AI models via Vertex AI while
maintaining OpenAI API compatibility.
"""Google Vertex AI LLM service with OpenAI API compatibility.
Provides access to Google's AI models via Vertex AI while maintaining
OpenAI API compatibility. Handles authentication using Google service
account credentials and constructs appropriate endpoint URLs for
different GCP regions and projects.
Reference:
https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/call-vertex-using-openai-library
https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/call-vertex-using-openai-library
"""
class InputParams(OpenAILLMService.InputParams):
"""Input parameters specific to Vertex AI."""
"""Input parameters specific to Vertex AI.
Parameters:
location: GCP region for Vertex AI endpoint (e.g., "us-east4").
project_id: Google Cloud project ID.
"""
# https://cloud.google.com/vertex-ai/generative-ai/docs/learn/locations
location: str = "us-east4"
@@ -58,11 +72,11 @@ class GoogleVertexLLMService(OpenAILLMService):
"""Initializes the VertexLLMService.
Args:
credentials (Optional[str]): JSON string of service account credentials.
credentials_path (Optional[str]): Path to the service account JSON file.
model (str): Model identifier. Defaults to "google/gemini-2.0-flash-001".
params (InputParams): Vertex AI input parameters.
**kwargs: Additional arguments for OpenAILLMService.
credentials: JSON string of service account credentials.
credentials_path: Path to the service account JSON file.
model: Model identifier (e.g., "google/gemini-2.0-flash-001").
params: Vertex AI input parameters including location and project.
**kwargs: Additional arguments passed to OpenAILLMService.
"""
params = params or OpenAILLMService.InputParams()
base_url = self._get_base_url(params)
@@ -74,7 +88,7 @@ class GoogleVertexLLMService(OpenAILLMService):
@staticmethod
def _get_base_url(params: InputParams) -> str:
"""Constructs the base URL for Vertex AI API."""
"""Construct the base URL for Vertex AI API."""
return (
f"https://{params.location}-aiplatform.googleapis.com/v1/"
f"projects/{params.project_id}/locations/{params.location}/endpoints/openapi"
@@ -82,14 +96,22 @@ class GoogleVertexLLMService(OpenAILLMService):
@staticmethod
def _get_api_token(credentials: Optional[str], credentials_path: Optional[str]) -> str:
"""Retrieves an authentication token using Google service account credentials.
"""Retrieve an authentication token using Google service account credentials.
Supports multiple authentication methods:
1. Direct JSON credentials string
2. Path to service account JSON file
3. Default application credentials (ADC)
Args:
credentials (Optional[str]): JSON string of service account credentials.
credentials_path (Optional[str]): Path to the service account JSON file.
credentials: JSON string of service account credentials.
credentials_path: Path to the service account JSON file.
Returns:
str: OAuth token for API authentication.
OAuth token for API authentication.
Raises:
ValueError: If no valid credentials are provided or found.
"""
creds: Optional[service_account.Credentials] = None

41
src/pipecat/services/google/rtvi.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google RTVI integration models and observer implementation.
This module provides integration with Google's services through the RTVI framework,
including models for search responses and an observer for handling Google-specific
frame types.
"""
from typing import List, Literal, Optional
from pydantic import BaseModel
@@ -16,22 +23,56 @@ from pipecat.services.google.frames import LLMSearchOrigin, LLMSearchResponseFra
class RTVISearchResponseMessageData(BaseModel):
"""Data payload for search response messages in RTVI protocol.
Parameters:
search_result: The search result text, if available.
rendered_content: The rendered content from the search, if available.
origins: List of search result origins with metadata.
"""
search_result: Optional[str]
rendered_content: Optional[str]
origins: List[LLMSearchOrigin]
class RTVIBotLLMSearchResponseMessage(BaseModel):
"""RTVI message for bot LLM search responses.
Parameters:
label: Always "rtvi-ai" for RTVI protocol messages.
type: Always "bot-llm-search-response" for this message type.
data: The search response data payload.
"""
label: Literal["rtvi-ai"] = "rtvi-ai"
type: Literal["bot-llm-search-response"] = "bot-llm-search-response"
data: RTVISearchResponseMessageData
class GoogleRTVIObserver(RTVIObserver):
"""RTVI observer for Google service integration.
Extends the base RTVIObserver to handle Google-specific frame types,
particularly LLM search response frames from Google services.
"""
def __init__(self, rtvi: RTVIProcessor):
"""Initialize the Google RTVI observer.
Args:
rtvi: The RTVI processor to send messages through.
"""
super().__init__(rtvi)
async def on_push_frame(self, data: FramePushed):
"""Process frames being pushed through the pipeline.
Handles Google-specific frames in addition to the base RTVI frame types.
Args:
data: Frame push event data containing frame and metadata.
"""
await super().on_push_frame(data)
frame = data.frame

133
src/pipecat/services/google/stt.py
View File

@@ -4,11 +4,19 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Google Cloud Speech-to-Text V2 service implementation for Pipecat.
This module provides a Google Cloud Speech-to-Text V2 service with streaming
support, enabling real-time speech recognition with features like automatic
punctuation, voice activity detection, and multi-language support.
"""
import asyncio
import json
import os
import time
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.tracing.service_decorators import traced_stt
# Suppress gRPC fork warnings
@@ -352,9 +360,15 @@ class GoogleSTTService(STTService):
Provides real-time speech recognition using Google Cloud's Speech-to-Text V2 API
with streaming support. Handles audio transcription and optional voice activity detection.
Implements automatic stream reconnection to handle Google's 4-minute streaming limit.
Attributes:
InputParams: Configuration parameters for the STT service.
STREAMING_LIMIT: Google Cloud's streaming limit in milliseconds (4 minutes).
Raises:
ValueError: If neither credentials nor credentials_path is provided.
ValueError: If project ID is not found in credentials.
"""
# Google Cloud's STT service has a connection time limit of 5 minutes per stream.
@@ -366,7 +380,7 @@ class GoogleSTTService(STTService):
class InputParams(BaseModel):
"""Configuration parameters for Google Speech-to-Text.
Attributes:
Parameters:
languages: Single language or list of recognition languages. First language is primary.
model: Speech recognition model to use.
use_separate_recognition_per_channel: Process each audio channel separately.
@@ -395,13 +409,25 @@ class GoogleSTTService(STTService):
@field_validator("languages", mode="before")
@classmethod
def validate_languages(cls, v) -> List[Language]:
"""Ensure languages is always a list.
Args:
v: Single Language enum or list of Language enums.
Returns:
List[Language]: List of configured languages.
"""
if isinstance(v, Language):
return [v]
return v
@property
def language_list(self) -> List[Language]:
"""Get languages as a guaranteed list."""
"""Get languages as a guaranteed list.
Returns:
List[Language]: List of configured languages.
"""
assert isinstance(self.languages, list)
return self.languages
@@ -424,10 +450,6 @@ class GoogleSTTService(STTService):
sample_rate: Audio sample rate in Hertz.
params: Configuration parameters for the service.
**kwargs: Additional arguments passed to STTService.
Raises:
ValueError: If neither credentials nor credentials_path is provided.
ValueError: If project ID is not found in credentials.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
@@ -436,7 +458,6 @@ class GoogleSTTService(STTService):
self._location = location
self._stream = None
self._config = None
self._request_queue = asyncio.Queue()
self._streaming_task = None
# Used for keep-alive logic
@@ -501,6 +522,11 @@ class GoogleSTTService(STTService):
}
def can_generate_metrics(self) -> bool:
"""Check if the service can generate metrics.
Returns:
bool: True, as this service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language | List[Language]) -> str | List[str]:
@@ -548,7 +574,11 @@ class GoogleSTTService(STTService):
await self._reconnect_if_needed()
async def set_model(self, model: str):
"""Update the service's recognition model."""
"""Update the service's recognition model.
Args:
model: The new recognition model to use.
"""
logger.debug(f"Switching STT model to: {model}")
await super().set_model(model)
self._settings["model"] = model
@@ -556,14 +586,29 @@ class GoogleSTTService(STTService):
await self._reconnect_if_needed()
async def start(self, frame: StartFrame):
"""Start the STT service and establish connection.
Args:
frame: The start frame triggering the service start.
"""
await super().start(frame)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the STT service and clean up resources.
Args:
frame: The end frame triggering the service stop.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the STT service and clean up resources.
Args:
frame: The cancel frame triggering the service cancellation.
"""
await super().cancel(frame)
await self._disconnect()
@@ -585,7 +630,7 @@ class GoogleSTTService(STTService):
"""Update service options dynamically.
Args:
languages: New list of recongition languages.
languages: New list of recognition languages.
model: New recognition model.
enable_automatic_punctuation: Enable/disable automatic punctuation.
enable_spoken_punctuation: Enable/disable spoken punctuation.
@@ -683,23 +728,15 @@ class GoogleSTTService(STTService):
),
)
self._request_queue = asyncio.Queue()
self._streaming_task = self.create_task(self._stream_audio())
async def _disconnect(self):
"""Clean up streaming recognition resources."""
if self._streaming_task:
logger.debug("Disconnecting from Google Speech-to-Text")
# Send sentinel value to stop request generator
await self._request_queue.put(None)
await self.cancel_task(self._streaming_task)
self._streaming_task = None
# Clear any remaining items in the queue
while not self._request_queue.empty():
try:
self._request_queue.get_nowait()
self._request_queue.task_done()
except asyncio.QueueEmpty:
break
async def _request_generator(self):
"""Generates requests for the streaming recognize method."""
@@ -714,29 +751,23 @@ class GoogleSTTService(STTService):
)
while True:
try:
audio_data = await self._request_queue.get()
if audio_data is None: # Sentinel value to stop
break
audio_data = await self._request_queue.get()
# Check streaming limit
if (int(time.time() * 1000) - self._stream_start_time) > self.STREAMING_LIMIT:
logger.debug("Streaming limit reached, initiating graceful reconnection")
# Instead of immediate reconnection, we'll break and let the stream close naturally
self._last_audio_input = self._audio_input
self._audio_input = []
self._restart_counter += 1
# Put the current audio chunk back in the queue
await self._request_queue.put(audio_data)
break
self._request_queue.task_done()
self._audio_input.append(audio_data)
yield cloud_speech.StreamingRecognizeRequest(audio=audio_data)
except asyncio.CancelledError:
# Check streaming limit
if (int(time.time() * 1000) - self._stream_start_time) > self.STREAMING_LIMIT:
logger.debug("Streaming limit reached, initiating graceful reconnection")
# Instead of immediate reconnection, we'll break and let the stream close naturally
self._last_audio_input = self._audio_input
self._audio_input = []
self._restart_counter += 1
# Put the current audio chunk back in the queue
await self._request_queue.put(audio_data)
break
finally:
self._request_queue.task_done()
self._audio_input.append(audio_data)
yield cloud_speech.StreamingRecognizeRequest(audio=audio_data)
except Exception as e:
logger.error(f"Error in request generator: {e}")
@@ -747,8 +778,6 @@ class GoogleSTTService(STTService):
try:
while True:
try:
self.start_watchdog()
if self._request_queue.empty():
# wait for 10ms in case we don't have audio
await asyncio.sleep(0.01)
@@ -763,8 +792,6 @@ class GoogleSTTService(STTService):
# Process responses
await self._process_responses(streaming_recognize)
self.reset_watchdog()
# If we're here, check if we need to reconnect
if (int(time.time() * 1000) - self._stream_start_time) > self.STREAMING_LIMIT:
logger.debug("Reconnecting stream after timeout")
@@ -779,15 +806,20 @@ class GoogleSTTService(STTService):
await asyncio.sleep(1) # Brief delay before reconnecting
self._stream_start_time = int(time.time() * 1000)
finally:
self.reset_watchdog()
except Exception as e:
logger.error(f"Error in streaming task: {e}")
await self.push_frame(ErrorFrame(str(e)))
async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]:
"""Process an audio chunk for STT transcription."""
"""Process an audio chunk for STT transcription.
Args:
audio: Raw audio bytes to transcribe.
Yields:
Frame: None (actual transcription frames are pushed via internal processing).
"""
if self._streaming_task:
# Queue the audio data
await self.start_ttfb_metrics()
@@ -804,17 +836,15 @@ class GoogleSTTService(STTService):
async def _process_responses(self, streaming_recognize):
"""Process streaming recognition responses."""
try:
async for response in streaming_recognize:
self.start_watchdog()
async for response in WatchdogAsyncIterator(
streaming_recognize, manager=self.task_manager
):
# Check streaming limit
if (int(time.time() * 1000) - self._stream_start_time) > self.STREAMING_LIMIT:
logger.debug("Stream timeout reached in response processing")
self.reset_watchdog()
break
if not response.results:
self.reset_watchdog()
continue
for result in response.results:
@@ -856,11 +886,8 @@ class GoogleSTTService(STTService):
result=result,
)
)
self.reset_watchdog()
except Exception as e:
logger.error(f"Error processing Google STT responses: {e}")
self.reset_watchdog()
# Re-raise the exception to let it propagate (e.g. in the case of a
# timeout, propagate to _stream_audio to reconnect)
raise

140
src/pipecat/services/google/tts.py
View File

@@ -4,7 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
import asyncio
"""Google Cloud Text-to-Speech service implementations.
This module provides integration with Google Cloud Text-to-Speech API,
offering both HTTP-based synthesis with SSML support and streaming synthesis
for real-time applications.
"""
import json
import os
@@ -43,6 +49,14 @@ except ModuleNotFoundError as e:
def language_to_google_tts_language(language: Language) -> Optional[str]:
"""Convert a Language enum to Google TTS language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding Google TTS language code, or None if not supported.
"""
language_map = {
# Afrikaans
Language.AF: "af-ZA",
@@ -203,7 +217,32 @@ def language_to_google_tts_language(language: Language) -> Optional[str]:
class GoogleHttpTTSService(TTSService):
"""Google Cloud Text-to-Speech HTTP service with SSML support.
Provides text-to-speech synthesis using Google Cloud's HTTP API with
comprehensive SSML support for voice customization, prosody control,
and styling options. Ideal for applications requiring fine-grained
control over speech output.
Note:
Requires Google Cloud credentials via service account JSON, credentials file,
or default application credentials (GOOGLE_APPLICATION_CREDENTIALS).
Chirp and Journey voices don't support SSML and will use plain text input.
"""
class InputParams(BaseModel):
"""Input parameters for Google HTTP TTS voice customization.
Parameters:
pitch: Voice pitch adjustment (e.g., "+2st", "-50%").
rate: Speaking rate adjustment (e.g., "slow", "fast", "125%").
volume: Volume adjustment (e.g., "loud", "soft", "+6dB").
emphasis: Emphasis level for the text.
language: Language for synthesis. Defaults to English.
gender: Voice gender preference.
google_style: Google-specific voice style.
"""
pitch: Optional[str] = None
rate: Optional[str] = None
volume: Optional[str] = None
@@ -222,6 +261,16 @@ class GoogleHttpTTSService(TTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initializes the Google HTTP TTS service.
Args:
credentials: JSON string containing Google Cloud service account credentials.
credentials_path: Path to Google Cloud service account JSON file.
voice_id: Google TTS voice identifier (e.g., "en-US-Standard-A").
sample_rate: Audio sample rate in Hz. If None, uses default.
params: Voice customization parameters including pitch, rate, volume, etc.
**kwargs: Additional arguments passed to parent TTSService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or GoogleHttpTTSService.InputParams()
@@ -245,11 +294,20 @@ class GoogleHttpTTSService(TTSService):
def _create_client(
self, credentials: Optional[str], credentials_path: Optional[str]
) -> texttospeech_v1.TextToSpeechAsyncClient:
"""Create authenticated Google Text-to-Speech client.
Args:
credentials: JSON string with service account credentials.
credentials_path: Path to service account JSON file.
Returns:
Authenticated TextToSpeechAsyncClient instance.
Raises:
ValueError: If no valid credentials are provided.
"""
creds: Optional[service_account.Credentials] = None
# Create a Google Cloud service account for the Cloud Text-to-Speech API
# Using either the provided credentials JSON string or the path to a service account JSON
# file, create a Google Cloud service account and use it to authenticate with the API.
if credentials:
# Use provided credentials JSON string
json_account_info = json.loads(credentials)
@@ -271,9 +329,22 @@ class GoogleHttpTTSService(TTSService):
return texttospeech_v1.TextToSpeechAsyncClient(credentials=creds)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Google HTTP TTS service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Google TTS language format.
Args:
language: The language to convert.
Returns:
The Google TTS-specific language code, or None if not supported.
"""
return language_to_google_tts_language(language)
def _construct_ssml(self, text: str) -> str:
@@ -324,6 +395,14 @@ class GoogleHttpTTSService(TTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Google's HTTP TTS API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:
@@ -381,19 +460,13 @@ class GoogleHttpTTSService(TTSService):
class GoogleTTSService(TTSService):
"""Text-to-Speech service using Google Cloud Text-to-Speech API.
"""Google Cloud Text-to-Speech streaming service.
Converts text to speech using Google's TTS models with streaming synthesis
for low latency. Supports multiple languages and voices.
Provides real-time text-to-speech synthesis using Google Cloud's streaming API
for low-latency applications. Optimized for Chirp 3 HD and Journey voices
with continuous audio streaming capabilities.
Args:
credentials: JSON string containing Google Cloud service account credentials.
credentials_path: Path to Google Cloud service account JSON file.
voice_id: Google TTS voice identifier (e.g., "en-US-Chirp3-HD-Charon").
sample_rate: Audio sample rate in Hz.
params: Language only.
Notes:
Note:
Requires Google Cloud credentials via service account JSON, file path, or
default application credentials (GOOGLE_APPLICATION_CREDENTIALS env var).
Only Chirp 3 HD and Journey voices are supported. Use GoogleHttpTTSService for other voices.
@@ -411,6 +484,12 @@ class GoogleTTSService(TTSService):
"""
class InputParams(BaseModel):
"""Input parameters for Google streaming TTS configuration.
Parameters:
language: Language for synthesis. Defaults to English.
"""
language: Optional[Language] = Language.EN
def __init__(
@@ -423,6 +502,16 @@ class GoogleTTSService(TTSService):
params: InputParams = InputParams(),
**kwargs,
):
"""Initializes the Google streaming TTS service.
Args:
credentials: JSON string containing Google Cloud service account credentials.
credentials_path: Path to Google Cloud service account JSON file.
voice_id: Google TTS voice identifier (e.g., "en-US-Chirp3-HD-Charon").
sample_rate: Audio sample rate in Hz. If None, uses default.
params: Language configuration parameters.
**kwargs: Additional arguments passed to parent TTSService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or GoogleTTSService.InputParams()
@@ -466,13 +555,34 @@ class GoogleTTSService(TTSService):
return texttospeech_v1.TextToSpeechAsyncClient(credentials=creds)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Google streaming TTS service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Google TTS language format.
Args:
language: The language to convert.
Returns:
The Google TTS-specific language code, or None if not supported.
"""
return language_to_google_tts_language(language)
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate streaming speech from text using Google's streaming API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech as it's generated.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:

78
src/pipecat/services/grok/llm.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Grok LLM service implementation using OpenAI-compatible interface.
This module provides a service for interacting with Grok's API through an
OpenAI-compatible interface, including specialized token usage tracking
and context aggregation functionality.
"""
from dataclasses import dataclass
from loguru import logger
@@ -23,13 +30,33 @@ from pipecat.services.openai.llm import (
@dataclass
class GrokContextAggregatorPair:
"""Pair of context aggregators for user and assistant interactions.
Provides a convenient container for managing both user and assistant
context aggregators together for Grok LLM interactions.
Parameters:
_user: The user context aggregator instance.
_assistant: The assistant context aggregator instance.
"""
_user: OpenAIUserContextAggregator
_assistant: OpenAIAssistantContextAggregator
def user(self) -> OpenAIUserContextAggregator:
"""Get the user context aggregator.
Returns:
The user context aggregator instance.
"""
return self._user
def assistant(self) -> OpenAIAssistantContextAggregator:
"""Get the assistant context aggregator.
Returns:
The assistant context aggregator instance.
"""
return self._assistant
@@ -38,12 +65,8 @@ class GrokLLMService(OpenAILLMService):
This service extends OpenAILLMService to connect to Grok's API endpoint while
maintaining full compatibility with OpenAI's interface and functionality.
Args:
api_key (str): The API key for accessing Grok's API
base_url (str, optional): The base URL for Grok API. Defaults to "https://api.x.ai/v1"
model (str, optional): The model identifier to use. Defaults to "grok-3-beta"
**kwargs: Additional keyword arguments passed to OpenAILLMService
Includes specialized token usage tracking that accumulates metrics during
processing and reports final totals.
"""
def __init__(
@@ -54,6 +77,14 @@ class GrokLLMService(OpenAILLMService):
model: str = "grok-3-beta",
**kwargs,
):
"""Initialize the GrokLLMService with API key and model.
Args:
api_key: The API key for accessing Grok's API.
base_url: The base URL for Grok API. Defaults to "https://api.x.ai/v1".
model: The model identifier to use. Defaults to "grok-3-beta".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
# Initialize counters for token usage metrics
self._prompt_tokens = 0
@@ -63,7 +94,16 @@ class GrokLLMService(OpenAILLMService):
self._is_processing = False
def create_client(self, api_key=None, base_url=None, **kwargs):
"""Create OpenAI-compatible client for Grok API endpoint."""
"""Create OpenAI-compatible client for Grok API endpoint.
Args:
api_key: The API key to use. If None, uses instance default.
base_url: The base URL to use. If None, uses instance default.
**kwargs: Additional arguments passed to client creation.
Returns:
The configured client instance for Grok API.
"""
logger.debug(f"Creating Grok client with api {base_url}")
return super().create_client(api_key, base_url, **kwargs)
@@ -75,8 +115,8 @@ class GrokLLMService(OpenAILLMService):
them once at the end of processing.
Args:
context (OpenAILLMContext): The context to process, containing messages
and other information needed for the LLM interaction.
context: The context to process, containing messages and other
information needed for the LLM interaction.
"""
# Reset all counters and flags at the start of processing
self._prompt_tokens = 0
@@ -107,8 +147,8 @@ class GrokLLMService(OpenAILLMService):
The final accumulated totals are reported at the end of processing.
Args:
tokens (LLMTokenUsage): The token usage metrics for the current chunk
of processing, containing prompt_tokens and completion_tokens counts.
tokens: The token usage metrics for the current chunk of processing,
containing prompt_tokens and completion_tokens counts.
"""
# Only accumulate metrics during active processing
if not self._is_processing:
@@ -130,22 +170,20 @@ class GrokLLMService(OpenAILLMService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> GrokContextAggregatorPair:
"""Create an instance of GrokContextAggregatorPair from an
OpenAILLMContext. Constructor keyword arguments for both the user and
assistant aggregators can be provided.
"""Create an instance of GrokContextAggregatorPair from an OpenAILLMContext.
Constructor keyword arguments for both the user and assistant aggregators
can be provided.
Args:
context (OpenAILLMContext): The LLM context.
user_params (LLMUserAggregatorParams, optional): User aggregator
parameters.
assistant_params (LLMAssistantAggregatorParams, optional): User
aggregator parameters.
context: The LLM context to create aggregators for.
user_params: Parameters for configuring the user aggregator.
assistant_params: Parameters for configuring the assistant aggregator.
Returns:
GrokContextAggregatorPair: A pair of context aggregators, one for
the user and one for the assistant, encapsulated in an
GrokContextAggregatorPair.
"""
context.set_llm_adapter(self.get_llm_adapter())

27
src/pipecat/services/groq/llm.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Groq LLM Service implementation using OpenAI-compatible interface."""
from loguru import logger
from pipecat.services.openai.llm import OpenAILLMService
@@ -14,12 +16,6 @@ class GroqLLMService(OpenAILLMService):
This service extends OpenAILLMService to connect to Groq's API endpoint while
maintaining full compatibility with OpenAI's interface and functionality.
Args:
api_key (str): The API key for accessing Groq's API
base_url (str, optional): The base URL for Groq API. Defaults to "https://api.groq.com/openai/v1"
model (str, optional): The model identifier to use. Defaults to "llama-3.3-70b-versatile"
**kwargs: Additional keyword arguments passed to OpenAILLMService
"""
def __init__(
@@ -30,9 +26,26 @@ class GroqLLMService(OpenAILLMService):
model: str = "llama-3.3-70b-versatile",
**kwargs,
):
"""Initialize Groq LLM service.
Args:
api_key: The API key for accessing Groq's API.
base_url: The base URL for Groq API. Defaults to "https://api.groq.com/openai/v1".
model: The model identifier to use. Defaults to "llama-3.3-70b-versatile".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
def create_client(self, api_key=None, base_url=None, **kwargs):
"""Create OpenAI-compatible client for Groq API endpoint."""
"""Create OpenAI-compatible client for Groq API endpoint.
Args:
api_key: API key for authentication. If None, uses instance api_key.
base_url: Base URL for the API. If None, uses instance base_url.
**kwargs: Additional arguments passed to the client constructor.
Returns:
An OpenAI-compatible client configured for Groq's API.
"""
logger.debug(f"Creating Groq client with api {base_url}")
return super().create_client(api_key, base_url, **kwargs)

22
src/pipecat/services/groq/stt.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Groq speech-to-text service implementation using Whisper models."""
from typing import Optional
from pipecat.services.whisper.base_stt import BaseWhisperSTTService, Transcription
@@ -15,15 +17,6 @@ class GroqSTTService(BaseWhisperSTTService):
Uses Groq's Whisper API to convert audio to text. Requires a Groq API key
set via the api_key parameter or GROQ_API_KEY environment variable.
Args:
model: Whisper model to use. Defaults to "whisper-large-v3-turbo".
api_key: Groq API key. Defaults to None.
base_url: API base URL. Defaults to "https://api.groq.com/openai/v1".
language: Language of the audio input. Defaults to English.
prompt: Optional text to guide the model's style or continue a previous segment.
temperature: Optional sampling temperature between 0 and 1. Defaults to 0.0.
**kwargs: Additional arguments passed to BaseWhisperSTTService.
"""
def __init__(
@@ -37,6 +30,17 @@ class GroqSTTService(BaseWhisperSTTService):
temperature: Optional[float] = None,
**kwargs,
):
"""Initialize Groq STT service.
Args:
model: Whisper model to use. Defaults to "whisper-large-v3-turbo".
api_key: Groq API key. Defaults to None.
base_url: API base URL. Defaults to "https://api.groq.com/openai/v1".
language: Language of the audio input. Defaults to English.
prompt: Optional text to guide the model's style or continue a previous segment.
temperature: Optional sampling temperature between 0 and 1. Defaults to 0.0.
**kwargs: Additional arguments passed to BaseWhisperSTTService.
"""
super().__init__(
model=model,
api_key=api_key,

40
src/pipecat/services/groq/tts.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Groq text-to-speech service implementation."""
import io
import wave
from typing import AsyncGenerator, Optional
@@ -25,7 +27,21 @@ except ModuleNotFoundError as e:
class GroqTTSService(TTSService):
"""Groq text-to-speech service implementation.
Provides text-to-speech synthesis using Groq's TTS API. The service
operates at a fixed 48kHz sample rate and supports various voices
and output formats.
"""
class InputParams(BaseModel):
"""Input parameters for Groq TTS configuration.
Parameters:
language: Language for speech synthesis. Defaults to English.
speed: Speech speed multiplier. Defaults to 1.0.
"""
language: Optional[Language] = Language.EN
speed: Optional[float] = 1.0
@@ -42,6 +58,17 @@ class GroqTTSService(TTSService):
sample_rate: Optional[int] = GROQ_SAMPLE_RATE,
**kwargs,
):
"""Initialize Groq TTS service.
Args:
api_key: Groq API key for authentication.
output_format: Audio output format. Defaults to "wav".
params: Additional input parameters for voice customization.
model_name: TTS model to use. Defaults to "playai-tts".
voice_id: Voice identifier to use. Defaults to "Celeste-PlayAI".
sample_rate: Audio sample rate. Must be 48000 Hz for Groq TTS.
**kwargs: Additional arguments passed to parent TTSService class.
"""
if sample_rate != self.GROQ_SAMPLE_RATE:
logger.warning(f"Groq TTS only supports {self.GROQ_SAMPLE_RATE}Hz sample rate. ")
@@ -71,10 +98,23 @@ class GroqTTSService(TTSService):
self._client = AsyncGroq(api_key=self._api_key)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Groq TTS service supports metrics generation.
"""
return True
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Groq's TTS API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech data.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
measuring_ttfb = True
await self.start_ttfb_metrics()

39
src/pipecat/services/image_service.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Image generation service implementation.
Provides base functionality for AI-powered image generation services that convert
text prompts into images.
"""
from abc import abstractmethod
from typing import AsyncGenerator
@@ -13,15 +19,48 @@ from pipecat.services.ai_service import AIService
class ImageGenService(AIService):
"""Base class for image generation services.
Processes TextFrames by using their content as prompts for image generation.
Subclasses must implement the run_image_gen method to provide actual image
generation functionality using their specific AI service.
"""
def __init__(self, **kwargs):
"""Initialize the image generation service.
Args:
**kwargs: Additional arguments passed to the parent AIService.
"""
super().__init__(**kwargs)
# Renders the image. Returns an Image object.
@abstractmethod
async def run_image_gen(self, prompt: str) -> AsyncGenerator[Frame, None]:
"""Generate an image from a text prompt.
This method must be implemented by subclasses to provide actual image
generation functionality using their specific AI service.
Args:
prompt: The text prompt to generate an image from.
Yields:
Frame: Frames containing the generated image (typically ImageRawFrame
or URLImageRawFrame).
"""
pass
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames for image generation.
TextFrames are used as prompts for image generation, while other frames
are passed through unchanged.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
if isinstance(frame, TextFrame):

184
src/pipecat/services/llm_service.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Base classes for Large Language Model services with function calling support."""
import asyncio
import inspect
from dataclasses import dataclass
@@ -41,23 +43,34 @@ FunctionCallHandler = Callable[["FunctionCallParams"], Awaitable[None]]
# Type alias for a callback function that handles the result of an LLM function call.
class FunctionCallResultCallback(Protocol):
"""Protocol for function call result callbacks.
Handles the result of an LLM function call execution.
"""
async def __call__(
self, result: Any, *, properties: Optional[FunctionCallResultProperties] = None
) -> None: ...
) -> None:
"""Call the result callback.
Args:
result: The result of the function call.
properties: Optional properties for the result.
"""
...
@dataclass
class FunctionCallParams:
"""Parameters for a function call.
Attributes:
function_name (str): The name of the function being called.
arguments (Mapping[str, Any]): The arguments for the function.
tool_call_id (str): A unique identifier for the function call.
llm (LLMService): The LLMService instance being used.
context (OpenAILLMContext): The LLM context.
result_callback (FunctionCallResultCallback): Callback to handle the result of the function call.
Parameters:
function_name: The name of the function being called.
tool_call_id: A unique identifier for the function call.
arguments: The arguments for the function.
llm: The LLMService instance being used.
context: The LLM context.
result_callback: Callback to handle the result of the function call.
"""
function_name: str
@@ -70,14 +83,14 @@ class FunctionCallParams:
@dataclass
class FunctionCallRegistryItem:
"""Represents an entry in our function call registry. This is what the user
registers.
"""Represents an entry in the function call registry.
Attributes:
function_name (Optional[str]): The name of the function.
handler (FunctionCallHandler): The handler for processing function call parameters.
cancel_on_interruption (bool): Flag indicating whether to cancel the call on interruption.
This is what the user registers when calling register_function.
Parameters:
function_name: The name of the function (None for catch-all handler).
handler: The handler for processing function call parameters.
cancel_on_interruption: Whether to cancel the call on interruption.
"""
function_name: Optional[str]
@@ -87,16 +100,17 @@ class FunctionCallRegistryItem:
@dataclass
class FunctionCallRunnerItem:
"""Represents an internal function call entry to our function call
runner. The runner executes function calls in order.
"""Internal function call entry for the function call runner.
Attributes:
registry_name (Optional[str]): The function call name registration (could be None).
function_name (str): The name of the function.
tool_call_id (str): A unique identifier for the function call.
arguments (Mapping[str, Any]): The arguments for the function.
context (OpenAILLMContext): The LLM context.
The runner executes function calls in order.
Parameters:
registry_item: The registry item containing handler information.
function_name: The name of the function.
tool_call_id: A unique identifier for the function call.
arguments: The arguments for the function.
context: The LLM context.
run_llm: Optional flag to control LLM execution after function call.
"""
registry_item: FunctionCallRegistryItem
@@ -108,22 +122,27 @@ class FunctionCallRunnerItem:
class LLMService(AIService):
"""This is the base class for all LLM services. It handles function calling
registration and execution. The class also provides event handlers.
"""Base class for all LLM services.
An event to know when an LLM service completion timeout occurs:
Handles function calling registration and execution with support for both
parallel and sequential execution modes. Provides event handlers for
completion timeouts and function call lifecycle events.
@task.event_handler("on_completion_timeout")
async def on_completion_timeout(service):
...
Event handlers:
on_completion_timeout: Called when an LLM completion timeout occurs.
on_function_calls_started: Called when function calls are received and
execution is about to start.
And an event to know that function calls have been received from the LLM
service and that we are going to start executing them:
@task.event_handler("on_function_calls_started")
async def on_function_calls_started(service, function_calls: Sequence[FunctionCallFromLLM]):
...
Example:
```python
@task.event_handler("on_completion_timeout")
async def on_completion_timeout(service):
logger.warning("LLM completion timed out")
@task.event_handler("on_function_calls_started")
async def on_function_calls_started(service, function_calls):
logger.info(f"Starting {len(function_calls)} function calls")
```
"""
# OpenAILLMAdapter is used as the default adapter since it aligns with most LLM implementations.
@@ -131,6 +150,13 @@ class LLMService(AIService):
adapter_class: Type[BaseLLMAdapter] = OpenAILLMAdapter
def __init__(self, run_in_parallel: bool = True, **kwargs):
"""Initialize the LLM service.
Args:
run_in_parallel: Whether to run function calls in parallel or sequentially.
Defaults to True.
**kwargs: Additional arguments passed to the parent AIService.
"""
super().__init__(**kwargs)
self._run_in_parallel = run_in_parallel
self._start_callbacks = {}
@@ -143,6 +169,11 @@ class LLMService(AIService):
self._register_event_handler("on_completion_timeout")
def get_llm_adapter(self) -> BaseLLMAdapter:
"""Get the LLM adapter instance.
Returns:
The adapter instance used for LLM communication.
"""
return self._adapter
def create_context_aggregator(
@@ -152,24 +183,57 @@ class LLMService(AIService):
user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
) -> Any:
"""Create a context aggregator for managing LLM conversation context.
Must be implemented by subclasses.
Args:
context: The LLM context to create an aggregator for.
user_params: Parameters for user message aggregation.
assistant_params: Parameters for assistant message aggregation.
Returns:
A context aggregator instance.
"""
pass
async def start(self, frame: StartFrame):
"""Start the LLM service.
Args:
frame: The start frame.
"""
await super().start(frame)
if not self._run_in_parallel:
await self._create_sequential_runner_task()
async def stop(self, frame: EndFrame):
"""Stop the LLM service.
Args:
frame: The end frame.
"""
await super().stop(frame)
if not self._run_in_parallel:
await self._cancel_sequential_runner_task()
async def cancel(self, frame: CancelFrame):
"""Cancel the LLM service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
if not self._run_in_parallel:
await self._cancel_sequential_runner_task()
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process a frame.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
if isinstance(frame, StartInterruptionFrame):
@@ -188,6 +252,18 @@ class LLMService(AIService):
*,
cancel_on_interruption: bool = True,
):
"""Register a function handler for LLM function calls.
Args:
function_name: The name of the function to handle. Use None to handle
all function calls with a catch-all handler.
handler: The function handler. Should accept a single FunctionCallParams
parameter.
start_callback: Legacy callback function (deprecated). Put initialization
code at the top of your handler instead.
cancel_on_interruption: Whether to cancel this function call when an
interruption occurs. Defaults to True.
"""
# Registering a function with the function_name set to None will run
# that handler for all functions
self._functions[function_name] = FunctionCallRegistryItem(
@@ -210,16 +286,38 @@ class LLMService(AIService):
self._start_callbacks[function_name] = start_callback
def unregister_function(self, function_name: Optional[str]):
"""Remove a registered function handler.
Args:
function_name: The name of the function handler to remove.
"""
del self._functions[function_name]
if self._start_callbacks[function_name]:
del self._start_callbacks[function_name]
def has_function(self, function_name: str):
"""Check if a function handler is registered.
Args:
function_name: The name of the function to check.
Returns:
True if the function is registered or if a catch-all handler (None)
is registered.
"""
if None in self._functions.keys():
return True
return function_name in self._functions.keys()
async def run_function_calls(self, function_calls: Sequence[FunctionCallFromLLM]):
"""Execute a sequence of function calls from the LLM.
Triggers the on_function_calls_started event and executes functions
either in parallel or sequentially based on the run_in_parallel setting.
Args:
function_calls: The function calls to execute.
"""
if len(function_calls) == 0:
return
@@ -257,7 +355,7 @@ class LLMService(AIService):
else:
await self._sequential_runner_queue.put(runner_item)
async def call_start_function(self, context: OpenAILLMContext, function_name: str):
async def _call_start_function(self, context: OpenAILLMContext, function_name: str):
if function_name in self._start_callbacks.keys():
await self._start_callbacks[function_name](function_name, self, context)
elif None in self._start_callbacks.keys():
@@ -272,6 +370,18 @@ class LLMService(AIService):
text_content: Optional[str] = None,
video_source: Optional[str] = None,
):
"""Request an image from a user.
Pushes a UserImageRequestFrame upstream to request an image from the
specified user.
Args:
user_id: The ID of the user to request an image from.
function_name: Optional function name associated with the request.
tool_call_id: Optional tool call ID associated with the request.
text_content: Optional text content/context for the image request.
video_source: Optional video source identifier.
"""
await self.push_frame(
UserImageRequestFrame(
user_id=user_id,
@@ -316,7 +426,7 @@ class LLMService(AIService):
)
# NOTE(aleix): This needs to be removed after we remove the deprecation.
await self.call_start_function(runner_item.context, runner_item.function_name)
await self._call_start_function(runner_item.context, runner_item.function_name)
# Push a function call in-progress downstream. This frame will let our
# assistant context aggregator know that we are in the middle of a

74
src/pipecat/services/lmnt/tts.py
View File

@@ -4,6 +4,8 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""LMNT text-to-speech service implementation."""
import json
from typing import AsyncGenerator, Optional
@@ -35,6 +37,14 @@ except ModuleNotFoundError as e:
def language_to_lmnt_language(language: Language) -> Optional[str]:
"""Convert a Language enum to LMNT language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding LMNT language code, or None if not supported.
"""
BASE_LANGUAGES = {
Language.DE: "de",
Language.EN: "en",
@@ -71,6 +81,13 @@ def language_to_lmnt_language(language: Language) -> Optional[str]:
class LmntTTSService(InterruptibleTTSService):
"""LMNT real-time text-to-speech service.
Provides real-time text-to-speech synthesis using LMNT's WebSocket API.
Supports streaming audio generation with configurable voice models and
language settings.
"""
def __init__(
self,
*,
@@ -81,6 +98,16 @@ class LmntTTSService(InterruptibleTTSService):
model: str = "aurora",
**kwargs,
):
"""Initialize the LMNT TTS service.
Args:
api_key: LMNT API key for authentication.
voice_id: ID of the voice to use for synthesis.
sample_rate: Audio sample rate. If None, uses default.
language: Language for synthesis. Defaults to English.
model: TTS model to use. Defaults to "aurora".
**kwargs: Additional arguments passed to parent InterruptibleTTSService.
"""
super().__init__(
push_stop_frames=True,
pause_frame_processing=True,
@@ -99,35 +126,71 @@ class LmntTTSService(InterruptibleTTSService):
self._receive_task = None
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as LMNT service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to LMNT service language format.
Args:
language: The language to convert.
Returns:
The LMNT-specific language code, or None if not supported.
"""
return language_to_lmnt_language(language)
async def start(self, frame: StartFrame):
"""Start the LMNT TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the LMNT TTS service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the LMNT TTS service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._disconnect()
async def push_frame(self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM):
"""Push a frame downstream with special handling for stop conditions.
Args:
frame: The frame to push.
direction: The direction to push the frame.
"""
await super().push_frame(frame, direction)
if isinstance(frame, (TTSStoppedFrame, StartInterruptionFrame)):
self._started = False
async def _connect(self):
"""Connect to LMNT WebSocket and start receive task."""
await self._connect_websocket()
if self._websocket and not self._receive_task:
self._receive_task = self.create_task(self._receive_task_handler(self._report_error))
async def _disconnect(self):
"""Disconnect from LMNT WebSocket and clean up tasks."""
if self._receive_task:
await self.cancel_task(self._receive_task)
self._receive_task = None
@@ -181,11 +244,13 @@ class LmntTTSService(InterruptibleTTSService):
self._websocket = None
def _get_websocket(self):
"""Get the WebSocket connection if available."""
if self._websocket:
return self._websocket
raise Exception("Websocket not connected")
async def flush_audio(self):
"""Flush any pending audio synthesis."""
if not self._websocket or self._websocket.closed:
return
await self._get_websocket().send(json.dumps({"flush": True}))
@@ -216,7 +281,14 @@ class LmntTTSService(InterruptibleTTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate TTS audio from text."""
"""Generate TTS audio from text using LMNT's streaming API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
try:

128
src/pipecat/services/mcp_service.py
View File

@@ -1,5 +1,13 @@
#
# Copyright (c) 2024–2025, Daily
#
# SPDX-License-Identifier: BSD 2-Clause License
#
"""MCP (Model Context Protocol) client for integrating external tools with LLMs."""
import json
from typing import Any, Dict, List, Optional, Union
from typing import Any, Dict, List, Tuple
from loguru import logger
@@ -9,9 +17,11 @@ from pipecat.utils.base_object import BaseObject
try:
from mcp import ClientSession, StdioServerParameters
from mcp.client.session_group import SseServerParameters
from mcp.client.session import ClientSession
from mcp.client.session_group import SseServerParameters, StreamableHttpParameters
from mcp.client.sse import sse_client
from mcp.client.stdio import stdio_client
from mcp.client.streamable_http import streamablehttp_client
except ModuleNotFoundError as e:
logger.error(f"Exception: {e}")
logger.error("In order to use an MCP client, you need to `pip install pipecat-ai[mcp]`.")
@@ -19,26 +29,57 @@ except ModuleNotFoundError as e:
class MCPClient(BaseObject):
"""Client for Model Context Protocol (MCP) servers.
Enables integration with MCP servers to provide external tools and resources
to LLMs. Supports both stdio and SSE server connections with automatic tool
registration and schema conversion.
Raises:
TypeError: If server_params is not a supported parameter type.
"""
def __init__(
self,
server_params: Union[StdioServerParameters, SseServerParameters],
server_params: Tuple[StdioServerParameters, SseServerParameters, StreamableHttpParameters],
**kwargs,
):
"""Initialize the MCP client with server parameters.
Args:
server_params: Server connection parameters (stdio or SSE).
**kwargs: Additional arguments passed to the parent BaseObject.
"""
super().__init__(**kwargs)
self._server_params = server_params
self._session = ClientSession
if isinstance(server_params, StdioServerParameters):
self._client = stdio_client
self._register_tools = self._stdio_register_tools
elif isinstance(server_params, SseServerParameters):
self._client = sse_client
self._register_tools = self._sse_register_tools
elif isinstance(server_params, StreamableHttpParameters):
self._client = streamablehttp_client
self._register_tools = self._streamable_http_register_tools
else:
raise TypeError(
f"{self} invalid argument type: `server_params` must be either StdioServerParameters or SseServerParameters."
f"{self} invalid argument type: `server_params` must be either StdioServerParameters, SseServerParameters, or StreamableHttpParameters."
)
async def register_tools(self, llm) -> ToolsSchema:
"""Register all available MCP tools with an LLM service.
Connects to the MCP server, discovers available tools, converts their
schemas to Pipecat format, and registers them with the LLM service.
Args:
llm: The Pipecat LLM service to register tools with.
Returns:
A ToolsSchema containing all successfully registered tools.
"""
tools_schema = await self._register_tools(llm)
return tools_schema
@@ -46,13 +87,13 @@ class MCPClient(BaseObject):
self, tool_name: str, tool_schema: Dict[str, Any]
) -> FunctionSchema:
"""Convert an mcp tool schema to Pipecat's FunctionSchema format.
Args:
tool_name: The name of the tool
tool_schema: The mcp tool schema
Returns:
A FunctionSchema instance
"""
logger.debug(f"Converting schema for tool '{tool_name}'")
logger.trace(f"Original schema: {json.dumps(tool_schema, indent=2)}")
@@ -71,7 +112,8 @@ class MCPClient(BaseObject):
return schema
async def _sse_register_tools(self, llm) -> ToolsSchema:
"""Register all available mcp.run tools with the LLM service.
"""Register all available mcp tools with the LLM service.
Args:
llm: The Pipecat LLM service to register tools with
Returns:
@@ -86,16 +128,11 @@ class MCPClient(BaseObject):
context: any,
result_callback: any,
) -> None:
"""Wrapper for mcp.run tool calls to match Pipecat's function call interface."""
"""Wrapper for mcp tool calls to match Pipecat's function call interface."""
logger.debug(f"Executing tool '{function_name}' with call ID: {tool_call_id}")
logger.trace(f"Tool arguments: {json.dumps(arguments, indent=2)}")
try:
async with self._client(
url=self._server_params.url,
headers=self._server_params.headers,
timeout=self._server_params.timeout,
sse_read_timeout=self._server_params.sse_read_timeout,
) as (read, write):
async with self._client(**self._server_params.model_dump()) as (read, write):
async with self._session(read, write) as session:
await session.initialize()
await self._call_tool(session, function_name, arguments, result_callback)
@@ -106,20 +143,17 @@ class MCPClient(BaseObject):
await result_callback(error_msg)
logger.debug(f"SSE server parameters: {self._server_params}")
logger.debug("Starting registration of mcp tools")
async with self._client(
url=self._server_params.url,
headers=self._server_params.headers,
timeout=self._server_params.timeout,
sse_read_timeout=self._server_params.sse_read_timeout,
) as (read, write):
async with self._client(**self._server_params.model_dump()) as (read, write):
async with self._session(read, write) as session:
await session.initialize()
tools_schema = await self._list_tools(session, mcp_tool_wrapper, llm)
return tools_schema
async def _stdio_register_tools(self, llm) -> ToolsSchema:
"""Register all available mcp.run tools with the LLM service.
"""Register all available mcp tools with the LLM service.
Args:
llm: The Pipecat LLM service to register tools with
Returns:
@@ -134,7 +168,7 @@ class MCPClient(BaseObject):
context: any,
result_callback: any,
) -> None:
"""Wrapper for mcp.run tool calls to match Pipecat's function call interface."""
"""Wrapper for mcp tool calls to match Pipecat's function call interface."""
logger.debug(f"Executing tool '{function_name}' with call ID: {tool_call_id}")
logger.trace(f"Tool arguments: {json.dumps(arguments, indent=2)}")
try:
@@ -148,7 +182,7 @@ class MCPClient(BaseObject):
logger.exception("Full exception details:")
await result_callback(error_msg)
logger.debug("Starting registration of mcp.run tools")
logger.debug("Starting registration of mcp tools")
async with self._client(self._server_params) as streams:
async with self._session(streams[0], streams[1]) as session:
@@ -156,6 +190,53 @@ class MCPClient(BaseObject):
tools_schema = await self._list_tools(session, mcp_tool_wrapper, llm)
return tools_schema
async def _streamable_http_register_tools(self, llm) -> ToolsSchema:
"""Register all available mcp tools with the LLM service using streamable HTTP.
Args:
llm: The Pipecat LLM service to register tools with
Returns:
A ToolsSchema containing all registered tools
"""
async def mcp_tool_wrapper(
function_name: str,
tool_call_id: str,
arguments: Dict[str, Any],
llm: any,
context: any,
result_callback: any,
) -> None:
"""Wrapper for mcp tool calls to match Pipecat's function call interface."""
logger.debug(f"Executing tool '{function_name}' with call ID: {tool_call_id}")
logger.trace(f"Tool arguments: {json.dumps(arguments, indent=2)}")
try:
async with self._client(**self._server_params.model_dump()) as (
read_stream,
write_stream,
_,
):
async with self._session(read_stream, write_stream) as session:
await session.initialize()
await self._call_tool(session, function_name, arguments, result_callback)
except Exception as e:
error_msg = f"Error calling mcp tool {function_name}: {str(e)}"
logger.error(error_msg)
logger.exception("Full exception details:")
await result_callback(error_msg)
logger.debug("Starting registration of mcp tools using streamable HTTP")
async with self._client(**self._server_params.model_dump()) as (
read_stream,
write_stream,
_,
):
async with self._session(read_stream, write_stream) as session:
await session.initialize()
tools_schema = await self._list_tools(session, mcp_tool_wrapper, llm)
return tools_schema
async def _call_tool(self, session, function_name, arguments, result_callback):
logger.debug(f"Calling mcp tool '{function_name}'")
try:
@@ -199,8 +280,7 @@ class MCPClient(BaseObject):
try:
# Convert the schema
function_schema = self._convert_mcp_schema_to_pipecat(
tool_name,
{"description": tool.description, "input_schema": tool.inputSchema},
tool_name, {"description": tool.description, "input_schema": tool.inputSchema}
)
# Register the wrapped function

51
src/pipecat/services/mem0/memory.py
View File

@@ -4,6 +4,13 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Mem0 memory service integration for Pipecat.
This module provides a memory service that integrates with Mem0 to store
and retrieve conversational memories, enhancing LLM context with relevant
historical information.
"""
from typing import Any, Dict, List, Optional
from loguru import logger
@@ -31,14 +38,21 @@ class Mem0MemoryService(FrameProcessor):
This service intercepts message frames in the pipeline, stores them in Mem0,
and enhances context with relevant memories before passing them downstream.
Args:
api_key (str): The API key for accessing Mem0's API
user_id (str): The user ID to associate with memories in Mem0
params (InputParams, optional): Configuration parameters for memory retrieval
Supports both local and cloud-based Mem0 configurations.
"""
class InputParams(BaseModel):
"""Configuration parameters for Mem0 memory service.
Parameters:
search_limit: Maximum number of memories to retrieve per query.
search_threshold: Minimum similarity threshold for memory retrieval.
api_version: API version to use for Mem0 client operations.
system_prompt: Prefix text for memory context messages.
add_as_system_message: Whether to add memories as system messages.
position: Position to insert memory messages in context.
"""
search_limit: int = Field(default=10, ge=1)
search_threshold: float = Field(default=0.1, ge=0.0, le=1.0)
api_version: str = Field(default="v2")
@@ -56,6 +70,19 @@ class Mem0MemoryService(FrameProcessor):
run_id: Optional[str] = None,
params: Optional[InputParams] = None,
):
"""Initialize the Mem0 memory service.
Args:
api_key: The API key for accessing Mem0's cloud API.
local_config: Local configuration for Mem0 client (alternative to cloud API).
user_id: The user ID to associate with memories in Mem0.
agent_id: The agent ID to associate with memories in Mem0.
run_id: The run ID to associate with memories in Mem0.
params: Configuration parameters for memory retrieval and storage.
Raises:
ValueError: If none of user_id, agent_id, or run_id are provided.
"""
# Important: Call the parent class __init__ first
super().__init__()
@@ -86,7 +113,7 @@ class Mem0MemoryService(FrameProcessor):
"""Store messages in Mem0.
Args:
messages: List of message dictionaries to store
messages: List of message dictionaries to store in memory.
"""
try:
logger.debug(f"Storing {len(messages)} messages in Mem0")
@@ -110,10 +137,10 @@ class Mem0MemoryService(FrameProcessor):
"""Retrieve relevant memories from Mem0.
Args:
query: The query to search for relevant memories
query: The query to search for relevant memories.
Returns:
List of relevant memory dictionaries
List of relevant memory dictionaries matching the query.
"""
try:
logger.debug(f"Retrieving memories for query: {query}")
@@ -154,8 +181,8 @@ class Mem0MemoryService(FrameProcessor):
"""Enhance the LLM context with relevant memories.
Args:
context: The OpenAILLMContext to enhance
query: The query to search for relevant memories
context: The OpenAILLMContext to enhance with memory information.
query: The query to search for relevant memories.
"""
# Skip if this is the same query we just processed
if self.last_query == query:
@@ -184,8 +211,8 @@ class Mem0MemoryService(FrameProcessor):
"""Process incoming frames, intercept context frames for memory integration.
Args:
frame: The incoming frame to process
direction: The direction of frame flow in the pipeline
frame: The incoming frame to process.
direction: The direction of frame flow in the pipeline.
"""
await super().process_frame(frame, direction)

81
src/pipecat/services/minimax/tts.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""MiniMax text-to-speech service implementation.
This module provides integration with MiniMax's T2A (Text-to-Audio) API
for streaming text-to-speech synthesis.
"""
import json
from typing import AsyncGenerator, Optional
@@ -25,6 +31,14 @@ from pipecat.utils.tracing.service_decorators import traced_tts
def language_to_minimax_language(language: Language) -> Optional[str]:
"""Convert a Language enum to MiniMax language format.
Args:
language: The Language enum value to convert.
Returns:
The corresponding MiniMax language name, or None if not supported.
"""
BASE_LANGUAGES = {
Language.AR: "Arabic",
Language.CS: "Czech",
@@ -71,24 +85,18 @@ def language_to_minimax_language(language: Language) -> Optional[str]:
class MiniMaxHttpTTSService(TTSService):
"""Text-to-speech service using MiniMax's T2A (Text-to-Audio) API.
Provides streaming text-to-speech synthesis using MiniMax's HTTP API
with support for various voice settings, emotions, and audio configurations.
Supports real-time audio streaming with configurable voice parameters.
Platform documentation:
https://www.minimax.io/platform/document/T2A%20V2?key=66719005a427f0c8a5701643
Args:
api_key: MiniMax API key for authentication.
group_id: MiniMax Group ID to identify project.
model: TTS model name (default: "speech-02-turbo"). Options include
"speech-02-hd", "speech-02-turbo", "speech-01-hd", "speech-01-turbo".
voice_id: Voice identifier (default: "Calm_Woman").
aiohttp_session: aiohttp.ClientSession for API communication.
sample_rate: Output audio sample rate in Hz (default: None, set from pipeline).
params: Additional configuration parameters.
"""
class InputParams(BaseModel):
"""Configuration parameters for MiniMax TTS.
Attributes:
Parameters:
language: Language for TTS generation.
speed: Speech speed (range: 0.5 to 2.0).
volume: Speech volume (range: 0 to 10).
@@ -117,6 +125,19 @@ class MiniMaxHttpTTSService(TTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the MiniMax TTS service.
Args:
api_key: MiniMax API key for authentication.
group_id: MiniMax Group ID to identify project.
model: TTS model name. Defaults to "speech-02-turbo". Options include
"speech-02-hd", "speech-02-turbo", "speech-01-hd", "speech-01-turbo".
voice_id: Voice identifier. Defaults to "Calm_Woman".
aiohttp_session: aiohttp.ClientSession for API communication.
sample_rate: Output audio sample rate in Hz. If None, uses pipeline default.
params: Additional configuration parameters.
**kwargs: Additional arguments passed to parent TTSService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or MiniMaxHttpTTSService.InputParams()
@@ -175,28 +196,62 @@ class MiniMaxHttpTTSService(TTSService):
self._settings["english_normalization"] = params.english_normalization
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as MiniMax service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to MiniMax service language format.
Args:
language: The language to convert.
Returns:
The MiniMax-specific language name, or None if not supported.
"""
return language_to_minimax_language(language)
def set_model_name(self, model: str):
"""Set the TTS model to use"""
"""Set the TTS model to use.
Args:
model: The model name to use for synthesis.
"""
self._model_name = model
def set_voice(self, voice: str):
"""Set the voice to use"""
"""Set the voice to use.
Args:
voice: The voice identifier to use for synthesis.
"""
self._voice_id = voice
if "voice_setting" in self._settings:
self._settings["voice_setting"]["voice_id"] = voice
async def start(self, frame: StartFrame):
"""Start the MiniMax TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
self._settings["audio_setting"]["sample_rate"] = self.sample_rate
logger.debug(f"MiniMax TTS initialized with sample rate: {self.sample_rate}")
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate TTS audio from text using MiniMax's streaming API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"{self}: Generating TTS [{text}]")
headers = {

48
src/pipecat/services/moondream/vision.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Moondream vision service implementation.
This module provides integration with the Moondream vision-language model
for image analysis and description generation.
"""
import asyncio
from typing import AsyncGenerator
@@ -23,7 +29,15 @@ except ModuleNotFoundError as e:
def detect_device():
"""Detects the appropriate device to run on, and return the device and dtype."""
"""Detect the appropriate device to run on.
Detects available hardware acceleration and selects the best device
and data type for optimal performance.
Returns:
tuple: A tuple containing (device, dtype) where device is a torch.device
and dtype is the recommended torch data type for that device.
"""
try:
import intel_extension_for_pytorch
@@ -40,9 +54,24 @@ def detect_device():
class MoondreamService(VisionService):
"""Moondream vision-language model service.
Provides image analysis and description generation using the Moondream
vision-language model. Supports various hardware acceleration options
including CUDA, MPS, and Intel XPU.
"""
def __init__(
self, *, model="vikhyatk/moondream2", revision="2024-08-26", use_cpu=False, **kwargs
):
"""Initialize the Moondream service.
Args:
model: Hugging Face model identifier for the Moondream model.
revision: Specific model revision to use.
use_cpu: Whether to force CPU usage instead of hardware acceleration.
**kwargs: Additional arguments passed to the parent VisionService.
"""
super().__init__(**kwargs)
self.set_model_name(model)
@@ -65,6 +94,15 @@ class MoondreamService(VisionService):
logger.debug("Loaded Moondream model")
async def run_vision(self, frame: VisionImageRawFrame) -> AsyncGenerator[Frame, None]:
"""Analyze an image and generate a description.
Args:
frame: Vision frame containing the image data and optional question text.
Yields:
Frame: TextFrame containing the generated image description, or ErrorFrame
if analysis fails.
"""
if not self._model:
logger.error(f"{self} error: Moondream model not available ({self.model_name})")
yield ErrorFrame("Moondream model not available")
@@ -73,6 +111,14 @@ class MoondreamService(VisionService):
logger.debug(f"Analyzing image: {frame}")
def get_image_description(frame: VisionImageRawFrame):
"""Generate description for the given image frame.
Args:
frame: Vision frame containing image data and question.
Returns:
str: Generated description of the image.
"""
image = Image.frombytes(frame.format, frame.size, frame.image)
image_embeds = self._model.encode_image(image)
description = self._model.answer_question(

165
src/pipecat/services/neuphonic/tts.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""Neuphonic text-to-speech service implementations.
This module provides WebSocket and HTTP-based integrations with Neuphonic's
text-to-speech API for real-time audio synthesis.
"""
import asyncio
import base64
import json
@@ -29,6 +35,7 @@ from pipecat.frames.frames import (
from pipecat.processors.frame_processor import FrameDirection
from pipecat.services.tts_service import InterruptibleTTSService, TTSService
from pipecat.transcriptions.language import Language
from pipecat.utils.asyncio.watchdog_async_iterator import WatchdogAsyncIterator
from pipecat.utils.tracing.service_decorators import traced_tts
try:
@@ -41,6 +48,14 @@ except ModuleNotFoundError as e:
def language_to_neuphonic_lang_code(language: Language) -> Optional[str]:
"""Convert a Language enum to Neuphonic language code.
Args:
language: The Language enum value to convert.
Returns:
The corresponding Neuphonic language code, or None if not supported.
"""
BASE_LANGUAGES = {
Language.DE: "de",
Language.EN: "en",
@@ -68,7 +83,21 @@ def language_to_neuphonic_lang_code(language: Language) -> Optional[str]:
class NeuphonicTTSService(InterruptibleTTSService):
"""Neuphonic real-time text-to-speech service using WebSocket streaming.
Provides real-time text-to-speech synthesis using Neuphonic's WebSocket API.
Supports interruption handling, keepalive connections, and configurable voice
parameters for high-quality speech generation.
"""
class InputParams(BaseModel):
"""Input parameters for Neuphonic TTS configuration.
Parameters:
language: Language for synthesis. Defaults to English.
speed: Speech speed multiplier. Defaults to 1.0.
"""
language: Optional[Language] = Language.EN
speed: Optional[float] = 1.0
@@ -83,6 +112,17 @@ class NeuphonicTTSService(InterruptibleTTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the Neuphonic TTS service.
Args:
api_key: Neuphonic API key for authentication.
voice_id: ID of the voice to use for synthesis.
url: WebSocket URL for the Neuphonic API.
sample_rate: Audio sample rate in Hz. Defaults to 22050.
encoding: Audio encoding format. Defaults to "pcm_linear".
params: Additional input parameters for TTS configuration.
**kwargs: Additional arguments passed to parent InterruptibleTTSService.
"""
super().__init__(
aggregate_sentences=True,
push_text_frames=False,
@@ -113,12 +153,26 @@ class NeuphonicTTSService(InterruptibleTTSService):
self._keepalive_task = None
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Neuphonic service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Neuphonic service language format.
Args:
language: The language to convert.
Returns:
The Neuphonic-specific language code, or None if not supported.
"""
return language_to_neuphonic_lang_code(language)
async def _update_settings(self, settings: Mapping[str, Any]):
"""Update service settings and reconnect with new configuration."""
if "voice_id" in settings:
self.set_voice(settings["voice_id"])
@@ -128,28 +182,56 @@ class NeuphonicTTSService(InterruptibleTTSService):
logger.info(f"Switching TTS to settings: [{self._settings}]")
async def start(self, frame: StartFrame):
"""Start the Neuphonic TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
await self._connect()
async def stop(self, frame: EndFrame):
"""Stop the Neuphonic TTS service.
Args:
frame: The end frame.
"""
await super().stop(frame)
await self._disconnect()
async def cancel(self, frame: CancelFrame):
"""Cancel the Neuphonic TTS service.
Args:
frame: The cancel frame.
"""
await super().cancel(frame)
await self._disconnect()
async def flush_audio(self):
"""Flush any pending audio synthesis by sending stop command."""
if self._websocket:
msg = {"text": "<STOP>"}
await self._websocket.send(json.dumps(msg))
async def push_frame(self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM):
"""Push a frame downstream with special handling for stop conditions.
Args:
frame: The frame to push.
direction: The direction to push the frame.
"""
await super().push_frame(frame, direction)
if isinstance(frame, (TTSStoppedFrame, StartInterruptionFrame)):
self._started = False
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process frames with special handling for speech control.
Args:
frame: The frame to process.
direction: The direction of frame processing.
"""
await super().process_frame(frame, direction)
# If we received a TTSSpeakFrame and the LLM response included text (it
@@ -163,6 +245,7 @@ class NeuphonicTTSService(InterruptibleTTSService):
await self.resume_processing_frames()
async def _connect(self):
"""Connect to Neuphonic WebSocket and start background tasks."""
await self._connect_websocket()
if self._websocket and not self._receive_task:
@@ -172,6 +255,7 @@ class NeuphonicTTSService(InterruptibleTTSService):
self._keepalive_task = self.create_task(self._keepalive_task_handler())
async def _disconnect(self):
"""Disconnect from Neuphonic WebSocket and clean up tasks."""
if self._receive_task:
await self.cancel_task(self._receive_task)
self._receive_task = None
@@ -183,6 +267,7 @@ class NeuphonicTTSService(InterruptibleTTSService):
await self._disconnect_websocket()
async def _connect_websocket(self):
"""Establish WebSocket connection to Neuphonic API."""
try:
if self._websocket and self._websocket.open:
return
@@ -208,6 +293,7 @@ class NeuphonicTTSService(InterruptibleTTSService):
await self._call_event_handler("on_connection_error", f"{e}")
async def _disconnect_websocket(self):
"""Close WebSocket connection and clean up state."""
try:
await self.stop_all_metrics()
@@ -221,7 +307,8 @@ class NeuphonicTTSService(InterruptibleTTSService):
self._websocket = None
async def _receive_messages(self):
async for message in self._websocket:
"""Receive and process messages from Neuphonic WebSocket."""
async for message in WatchdogAsyncIterator(self._websocket, manager=self.task_manager):
if isinstance(message, str):
msg = json.loads(message)
if msg.get("data", {}).get("audio") is not None:
@@ -232,11 +319,15 @@ class NeuphonicTTSService(InterruptibleTTSService):
await self.push_frame(frame)
async def _keepalive_task_handler(self):
"""Handle keepalive messages to maintain WebSocket connection."""
KEEPALIVE_SLEEP = 10 if self.task_manager.task_watchdog_enabled else 3
while True:
await asyncio.sleep(10)
self.reset_watchdog()
await asyncio.sleep(KEEPALIVE_SLEEP)
await self._send_text("")
async def _send_text(self, text: str):
"""Send text to Neuphonic WebSocket for synthesis."""
if self._websocket:
msg = {"text": text}
logger.debug(f"Sending text to websocket: {msg}")
@@ -244,6 +335,14 @@ class NeuphonicTTSService(InterruptibleTTSService):
@traced_tts
async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
"""Generate speech from text using Neuphonic's streaming API.
Args:
text: The text to synthesize into speech.
Yields:
Frame: Audio frames containing the synthesized speech.
"""
logger.debug(f"Generating TTS: [{text}]")
try:
@@ -271,19 +370,21 @@ class NeuphonicTTSService(InterruptibleTTSService):
class NeuphonicHttpTTSService(TTSService):
"""Neuphonic Text-to-Speech service using HTTP streaming.
"""Neuphonic text-to-speech service using HTTP streaming.
Args:
api_key: Neuphonic API key
voice_id: ID of the voice to use
url: Base URL for the Neuphonic API (default: "https://api.neuphonic.com")
sample_rate: Sample rate for audio output (default: 22050Hz)
encoding: Audio encoding format (default: "pcm_linear")
params: Additional parameters for TTS generation including language and speed
**kwargs: Additional keyword arguments passed to the parent class
Provides text-to-speech synthesis using Neuphonic's HTTP API with server-sent
events for streaming audio delivery. Suitable for applications that prefer
HTTP-based communication over WebSocket connections.
"""
class InputParams(BaseModel):
"""Input parameters for Neuphonic HTTP TTS configuration.
Parameters:
language: Language for synthesis. Defaults to English.
speed: Speech speed multiplier. Defaults to 1.0.
"""
language: Optional[Language] = Language.EN
speed: Optional[float] = 1.0
@@ -298,6 +399,17 @@ class NeuphonicHttpTTSService(TTSService):
params: Optional[InputParams] = None,
**kwargs,
):
"""Initialize the Neuphonic HTTP TTS service.
Args:
api_key: Neuphonic API key for authentication.
voice_id: ID of the voice to use for synthesis.
url: Base URL for the Neuphonic HTTP API.
sample_rate: Audio sample rate in Hz. Defaults to 22050.
encoding: Audio encoding format. Defaults to "pcm_linear".
params: Additional input parameters for TTS configuration.
**kwargs: Additional arguments passed to parent TTSService.
"""
super().__init__(sample_rate=sample_rate, **kwargs)
params = params or NeuphonicHttpTTSService.InputParams()
@@ -313,12 +425,38 @@ class NeuphonicHttpTTSService(TTSService):
self.set_voice(voice_id)
def can_generate_metrics(self) -> bool:
"""Check if this service can generate processing metrics.
Returns:
True, as Neuphonic HTTP service supports metrics generation.
"""
return True
def language_to_service_language(self, language: Language) -> Optional[str]:
"""Convert a Language enum to Neuphonic service language format.
Args:
language: The language to convert.
Returns:
The Neuphonic-specific language code, or None if not supported.
"""
return language_to_neuphonic_lang_code(language)
async def start(self, frame: StartFrame):
"""Start the Neuphonic HTTP TTS service.
Args:
frame: The start frame containing initialization parameters.
"""
await super().start(frame)
async def flush_audio(self):
"""Flush any pending audio synthesis.
Note:
HTTP-based service doesn't require explicit flushing.
"""
pass
@traced_tts
@@ -326,9 +464,10 @@ class NeuphonicHttpTTSService(TTSService):
"""Generate speech from text using Neuphonic streaming API.
Args:
text: The text to convert to speech
text: The text to convert to speech.
Yields:
Frames containing audio data and status information
Frame: Audio frames containing the synthesized speech and status information.
"""
logger.debug(f"Generating TTS: [{text}]")

28
src/pipecat/services/nim/llm.py
View File

@@ -4,6 +4,12 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""NVIDIA NIM API service implementation.
This module provides a service for interacting with NVIDIA's NIM (NVIDIA Inference
Microservice) API while maintaining compatibility with the OpenAI-style interface.
"""
from pipecat.metrics.metrics import LLMTokenUsage
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContext
from pipecat.services.openai.llm import OpenAILLMService
@@ -15,12 +21,6 @@ class NimLLMService(OpenAILLMService):
This service extends OpenAILLMService to work with NVIDIA's NIM API while maintaining
compatibility with the OpenAI-style interface. It specifically handles the difference
in token usage reporting between NIM (incremental) and OpenAI (final summary).
Args:
api_key (str): The API key for accessing NVIDIA's NIM API
base_url (str, optional): The base URL for NIM API. Defaults to "https://integrate.api.nvidia.com/v1"
model (str, optional): The model identifier to use. Defaults to "nvidia/llama-3.1-nemotron-70b-instruct"
**kwargs: Additional keyword arguments passed to OpenAILLMService
"""
def __init__(
@@ -31,6 +31,14 @@ class NimLLMService(OpenAILLMService):
model: str = "nvidia/llama-3.1-nemotron-70b-instruct",
**kwargs,
):
"""Initialize the NimLLMService.
Args:
api_key: The API key for accessing NVIDIA's NIM API.
base_url: The base URL for NIM API. Defaults to "https://integrate.api.nvidia.com/v1".
model: The model identifier to use. Defaults to "nvidia/llama-3.1-nemotron-70b-instruct".
**kwargs: Additional keyword arguments passed to OpenAILLMService.
"""
super().__init__(api_key=api_key, base_url=base_url, model=model, **kwargs)
# Counters for accumulating token usage metrics
self._prompt_tokens = 0
@@ -47,8 +55,8 @@ class NimLLMService(OpenAILLMService):
them once at the end of processing.
Args:
context (OpenAILLMContext): The context to process, containing messages
and other information needed for the LLM interaction.
context: The context to process, containing messages and other information
needed for the LLM interaction.
"""
# Reset all counters and flags at the start of processing
self._prompt_tokens = 0
@@ -79,8 +87,8 @@ class NimLLMService(OpenAILLMService):
The final accumulated totals are reported at the end of processing.
Args:
tokens (LLMTokenUsage): The token usage metrics for the current chunk
of processing, containing prompt_tokens and completion_tokens counts.
tokens: The token usage metrics for the current chunk of processing,
containing prompt_tokens and completion_tokens counts.
"""
# Only accumulate metrics during active processing
if not self._is_processing:

15
src/pipecat/services/ollama/llm.py
View File

@@ -4,9 +4,24 @@
# SPDX-License-Identifier: BSD 2-Clause License
#
"""OLLama LLM service implementation for Pipecat AI framework."""
from pipecat.services.openai.llm import OpenAILLMService
class OLLamaLLMService(OpenAILLMService):
"""OLLama LLM service that provides local language model capabilities.
This service extends OpenAILLMService to work with locally hosted OLLama models,
providing a compatible interface for running large language models locally.
"""
def __init__(self, *, model: str = "llama2", base_url: str = "http://localhost:11434/v1"):
"""Initialize OLLama LLM service.
Args:
model: The OLLama model to use. Defaults to "llama2".
base_url: The base URL for the OLLama API endpoint.
Defaults to "http://localhost:11434/v1".
"""
super().__init__(model=model, base_url=base_url, api_key="ollama")

Some files were not shown because too many files have changed in this diff Show More