Merge pull request #1413 from pipecat-ai/aleix/llm-user-aggregator-emulate-fixes

LLMUserContextAggregator: fix emulated user started/stopped speaking issues

CHANGELOG.md

@@ -50,6 +50,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Added new `sample_rate` constructor parameter to `TavusVideoService` to allow
   changing the output sample rate.
 
+- Added new `NeuphonicTTSService`.
+  (see https://neuphonic.com)
+
 - Added new `UltravoxSTTService`.
   (see https://github.com/fixie-ai/ultravox)
 
@@ -269,6 +272,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Add foundational example `07w-interruptible-fal.py`, showing `FalSTTService`.
 
+- Added a new Ultravox example
+  `examples/foundational/07u-interruptible-ultravox.py`.
+
+- Added new Neuphonic examples
+  `examples/foundational/07v-interruptible-neuphonic.py` and
+  `examples/foundational/07v-interruptible-neuphonic-http.py`.
+
 - Added a new example `examples/foundational/36-user-email-gathering.py` to show
   how to gather user emails. The example uses's Cartesia's `<spell></spell>`
   tags and Rime `spell()` function to spell out the emails for confirmation.
@@ -367,6 +377,9 @@ stt = DeepgramSTTService(..., live_options=LiveOptions(model="nova-2-general"))
 
 ### Fixed
 
+- Fixed an issue that would cause undesired interruptions via
+  `EmulateUserStartedSpeakingFrame`.
+
 - Fixed a `GoogleLLMService` that was causing an exception when sending inline
   audio in some cases.
 
@@ -383,10 +396,6 @@ stt = DeepgramSTTService(..., live_options=LiveOptions(model="nova-2-general"))
 
 - Fixed `match_endofsentence` support for ellipses.
 
-- Fixed an issue that would cause undesired interruptions via
-  `EmulateUserStartedSpeakingFrame` when only interim transcriptions (i.e. no
-  final transcriptions) where received.
-
 - Fixed an issue where `EndTaskFrame` was not triggering
   `on_client_disconnected` or closing the WebSocket in FastAPI.
 

pyproject.toml

@@ -42,7 +42,7 @@ Website = "https://pipecat.ai"
 anthropic = [ "anthropic~=0.49.0" ]
 assemblyai = [ "assemblyai~=0.37.0" ]
 aws = [ "boto3~=1.37.16" ]
-azure = [ "azure-cognitiveservices-speech~=1.43.0"]
+azure = [ "azure-cognitiveservices-speech~=1.42.0"]
 canonical = [ "aiofiles~=24.1.0" ]
 cartesia = [ "cartesia~=1.4.0", "websockets~=13.1" ]
 neuphonic = [ "pyneuphonic~=1.5.13", "websockets~=13.1" ]

src/pipecat/processors/aggregators/llm_response.py

@@ -5,7 +5,6 @@
 #
 
 import asyncio
-import time
 from abc import abstractmethod
 from typing import Dict, List
 
@@ -222,17 +221,15 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
         self,
         context: OpenAILLMContext,
         aggregation_timeout: float = 1.0,
-        bot_interruption_timeout: float = 5.0,
         **kwargs,
     ):
         super().__init__(context=context, role="user", **kwargs)
         self._aggregation_timeout = aggregation_timeout
-        self._bot_interruption_timeout = bot_interruption_timeout
 
         self._seen_interim_results = False
         self._user_speaking = False
-        self._last_user_speaking_time = 0
         self._emulating_vad = False
+        self._waiting_for_aggregation = False
 
         self._aggregation_event = asyncio.Event()
         self._aggregation_task = None
@@ -240,6 +237,7 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
     def reset(self):
         super().reset()
         self._seen_interim_results = False
+        self._waiting_for_aggregation = False
 
     async def handle_aggregation(self, aggregation: str):
         self._context.add_message({"role": self.role, "content": self._aggregation})
@@ -285,14 +283,11 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
 
             # Reset the aggregation. Reset it before pushing it down, otherwise
             # if the tasks gets cancelled we won't be able to clear things up.
-            self._aggregation = ""
+            self.reset()
 
             frame = OpenAILLMContextFrame(self._context)
             await self.push_frame(frame)
 
-            # Reset our accumulator state.
-            self.reset()
-
     async def _start(self, frame: StartFrame):
         self._create_aggregation_task()
 
@@ -303,12 +298,14 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
         await self._cancel_aggregation_task()
 
     async def _handle_user_started_speaking(self, _: UserStartedSpeakingFrame):
-        self._last_user_speaking_time = time.time()
         self._user_speaking = True
+        self._waiting_for_aggregation = True
 
     async def _handle_user_stopped_speaking(self, _: UserStoppedSpeakingFrame):
-        self._last_user_speaking_time = time.time()
         self._user_speaking = False
+        # We just stopped speaking. Let's see if there's some aggregation to
+        # push. If the last thing we saw is an interim transcription, let's wait
+        # pushing the aggregation as we will probably get a final transcription.
         if not self._seen_interim_results:
             await self.push_aggregation()
 
@@ -361,18 +358,13 @@ class LLMUserContextAggregator(LLMContextResponseAggregator):
         frame we might want to interrupt the bot.
 
         """
-        if not self._user_speaking:
-            diff_time = time.time() - self._last_user_speaking_time
-            if diff_time > self._bot_interruption_timeout:
-                # If we reach this case we received a transcription but VAD was
-                # not able to detect voice (e.g. when you whisper a short
-                # utterance). So, we need to emulate VAD (i.e. user
-                # start/stopped speaking).
-                await self.push_frame(EmulateUserStartedSpeakingFrame(), FrameDirection.UPSTREAM)
-                self._emulating_vad = True
-
-                # Reset time so we don't interrupt again right away.
-                self._last_user_speaking_time = time.time()
+        if not self._user_speaking and not self._waiting_for_aggregation:
+            # If we reach this case we received a transcription but VAD was not
+            # able to detect voice (e.g. when you whisper a short
+            # utterance). So, we need to emulate VAD (i.e. user start/stopped
+            # speaking).
+            await self.push_frame(EmulateUserStartedSpeakingFrame(), FrameDirection.UPSTREAM)
+            self._emulating_vad = True
 
 
 class LLMAssistantContextAggregator(LLMContextResponseAggregator):
@@ -554,14 +546,11 @@ class LLMUserResponseAggregator(LLMUserContextAggregator):
 
             # Reset the aggregation. Reset it before pushing it down, otherwise
             # if the tasks gets cancelled we won't be able to clear things up.
-            self._aggregation = ""
+            self.reset()
 
             frame = LLMMessagesFrame(self._context.messages)
             await self.push_frame(frame)
 
-            # Reset our accumulator state.
-            self.reset()
-
 
 class LLMAssistantResponseAggregator(LLMAssistantContextAggregator):
     def __init__(self, messages: List[dict] = [], **kwargs):
@@ -573,10 +562,7 @@ class LLMAssistantResponseAggregator(LLMAssistantContextAggregator):
 
             # Reset the aggregation. Reset it before pushing it down, otherwise
             # if the tasks gets cancelled we won't be able to clear things up.
-            self._aggregation = ""
+            self.reset()
 
             frame = LLMMessagesFrame(self._context.messages)
             await self.push_frame(frame)
-
-            # Reset our accumulator state.
-            self.reset()

src/pipecat/transports/base_input.py

@@ -152,6 +152,7 @@ class BaseInputTransport(FrameProcessor):
     async def _handle_user_interruption(self, frame: Frame):
         if isinstance(frame, UserStartedSpeakingFrame):
             logger.debug("User started speaking")
+            await self.push_frame(frame)
             # Make sure we notify about interruptions quickly out-of-band.
             if self.interruptions_allowed:
                 await self._start_interruption()
@@ -161,12 +162,11 @@ class BaseInputTransport(FrameProcessor):
                 await self.push_frame(StartInterruptionFrame())
         elif isinstance(frame, UserStoppedSpeakingFrame):
             logger.debug("User stopped speaking")
+            await self.push_frame(frame)
             if self.interruptions_allowed:
                 await self._stop_interruption()
                 await self.push_frame(StopInterruptionFrame())
 
-        await self.push_frame(frame)
-
     #
     # Audio input
     #

src/pipecat/transports/services/livekit.py

@@ -599,13 +599,6 @@ class LiveKitTransport(BaseTransport):
             )
             await self._output.send_message(frame)
 
-    async def cleanup(self):
-        if self._input:
-            await self._input.cleanup()
-        if self._output:
-            await self._output.cleanup()
-        await self._client.disconnect()
-
     async def on_room_event(self, event):
         # Handle room events
         pass

tests/test_context_aggregators.py

@@ -44,8 +44,6 @@ from pipecat.tests.utils import SleepFrame, run_test
 
 AGGREGATION_TIMEOUT = 0.1
 AGGREGATION_SLEEP = 0.15
-BOT_INTERRUPTION_TIMEOUT = 0.2
-BOT_INTERRUPTION_SLEEP = 0.25
 
 
 class BaseTestUserContextAggregator:
@@ -388,14 +386,13 @@ class BaseTestUserContextAggregator:
         aggregator = self.AGGREGATOR_CLASS(
             context,
             aggregation_timeout=AGGREGATION_TIMEOUT,
-            bot_interruption_timeout=BOT_INTERRUPTION_TIMEOUT,
         )
         frames_to_send = [
             UserStartedSpeakingFrame(),
             InterimTranscriptionFrame(text="How ", user_id="cat", timestamp=""),
             SleepFrame(),
             UserStoppedSpeakingFrame(),
-            SleepFrame(BOT_INTERRUPTION_SLEEP),
+            SleepFrame(AGGREGATION_SLEEP),
             InterimTranscriptionFrame(text="are you?", user_id="cat", timestamp=""),
             TranscriptionFrame(text="How are you?", user_id="cat", timestamp=""),
             SleepFrame(sleep=AGGREGATION_SLEEP),
@@ -405,12 +402,10 @@ class BaseTestUserContextAggregator:
             UserStoppedSpeakingFrame,
             *self.EXPECTED_CONTEXT_FRAMES,
         ]
-        expected_up_frames = [EmulateUserStartedSpeakingFrame, EmulateUserStoppedSpeakingFrame]
         await run_test(
             aggregator,
             frames_to_send=frames_to_send,
             expected_down_frames=expected_down_frames,
-            expected_up_frames=expected_up_frames,
         )
         self.check_message_content(context, 0, "How are you?")