Additional LLM and TTS metrics (#343)

* added llm and tts usage metrics

* Metrics debug logging

* cleanup

examples/foundational/06-listen-and-respond.py

@@ -9,14 +9,15 @@ import aiohttp
 import os
 import sys
 
-from pipecat.frames.frames import LLMMessagesFrame
+from pipecat.frames.frames import Frame, LLMMessagesFrame, MetricsFrame
 from pipecat.pipeline.pipeline import Pipeline
 from pipecat.pipeline.runner import PipelineRunner
-from pipecat.pipeline.task import PipelineTask
+from pipecat.pipeline.task import PipelineParams, PipelineTask
 from pipecat.processors.aggregators.llm_response import (
     LLMAssistantResponseAggregator,
     LLMUserResponseAggregator,
 )
+from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
 from pipecat.processors.logger import FrameLogger
 from pipecat.services.elevenlabs import ElevenLabsTTSService
 from pipecat.services.openai import OpenAILLMService
@@ -34,6 +35,14 @@ logger.remove(0)
 logger.add(sys.stderr, level="DEBUG")
 
 
+class MetricsLogger(FrameProcessor):
+    async def process_frame(self, frame: Frame, direction: FrameDirection):
+        if isinstance(frame, MetricsFrame):
+            print(
+                f"!!! MetricsFrame: {frame}, ttfb: {frame.ttfb}, processing: {frame.processing}, tokens: {frame.tokens}, characters: {frame.characters}")
+        await self.push_frame(frame, direction)
+
+
 async def main():
     async with aiohttp.ClientSession() as session:
         (room_url, token) = await configure(session)
@@ -58,11 +67,10 @@ async def main():
 
         llm = OpenAILLMService(
             api_key=os.getenv("OPENAI_API_KEY"),
-            model="gpt-4o")
+            model="gpt-4o"
+        )
 
-        fl = FrameLogger("!!! after LLM", "red")
+        ml = MetricsLogger()
-        fltts = FrameLogger("@@@ out of tts", "green")
-        flend = FrameLogger("### out of the end", "magenta")
 
         messages = [
             {
@@ -77,15 +85,18 @@ async def main():
             transport.input(),
             tma_in,
             llm,
-            fl,
             tts,
-            fltts,
+            ml,
             transport.output(),
             tma_out,
-            flend
         ])
 
         task = PipelineTask(pipeline)
+        task = PipelineTask(pipeline, PipelineParams(
+            allow_interruptions=True,
+            enable_metrics=True,
+            report_only_initial_ttfb=False,
+        ))
 
         @transport.event_handler("on_first_participant_joined")
         async def on_first_participant_joined(transport, participant):

src/pipecat/frames/frames.py

@@ -276,6 +276,8 @@ class MetricsFrame(SystemFrame):
     """
     ttfb: List[Mapping[str, Any]] | None = None
     processing: List[Mapping[str, Any]] | None = None
+    tokens: List[Mapping[str, Any]] | None = None
+    characters: List[Mapping[str, Any]] | None = None
 
 #
 # Control frames

src/pipecat/processors/frameworks/rtvi.py

@@ -537,7 +537,11 @@ class RTVIProcessor(FrameProcessor):
             processors.extend(pipeline.processors_with_metrics())
             ttfb = [{"processor": p.name, "value": 0.0} for p in processors]
             processing = [{"processor": p.name, "value": 0.0} for p in processors]
-            await self.push_frame(MetricsFrame(ttfb=ttfb, processing=processing))
+            tokens = [{"processor": p.name, "value": {"prompt_tokens": 0,
+                                                      "completion_tokens": 0,
+                                                      "total_tokens": 0}} for p in processors]
+            characters = [{"processor": p.name, "value": 0} for p in processors]
+            await self.push_frame(MetricsFrame(ttfb=ttfb, processing=processing, tokens=tokens, characters=characters))
 
         self._pipeline = pipeline
 

src/pipecat/services/azure.py

@@ -18,6 +18,7 @@ from pipecat.frames.frames import (
     EndFrame,
     ErrorFrame,
     Frame,
+    MetricsFrame,
     StartFrame,
     SystemFrame,
     TranscriptionFrame,
@@ -87,7 +88,13 @@ class AzureTTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
-
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
         await self.start_ttfb_metrics()
 
         ssml = (

src/pipecat/services/cartesia.py

@@ -21,6 +21,7 @@ from pipecat.frames.frames import (
     StartFrame,
     EndFrame,
     TextFrame,
+    MetricsFrame,
     LLMFullResponseEndFrame
 )
 from pipecat.services.ai_services import TTSService
@@ -200,6 +201,13 @@ class CartesiaTTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
 
         try:
             if not self._websocket:

src/pipecat/services/deepgram.py

@@ -15,6 +15,7 @@ from pipecat.frames.frames import (
     ErrorFrame,
     Frame,
     InterimTranscriptionFrame,
+    MetricsFrame,
     StartFrame,
     SystemFrame,
     TranscriptionFrame)
@@ -70,7 +71,13 @@ class DeepgramTTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
-
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
         base_url = self._base_url
         request_url = f"{base_url}?model={self._voice}&encoding={self._encoding}&container=none&sample_rate={self._sample_rate}"
         headers = {"authorization": f"token {self._api_key}"}

src/pipecat/services/elevenlabs.py

@@ -8,7 +8,7 @@ import aiohttp
 
 from typing import AsyncGenerator
 
-from pipecat.frames.frames import AudioRawFrame, ErrorFrame, Frame
+from pipecat.frames.frames import AudioRawFrame, ErrorFrame, Frame, MetricsFrame
 from pipecat.services.ai_services import TTSService
 
 from loguru import logger
@@ -40,7 +40,13 @@ class ElevenLabsTTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
-
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
         url = f"https://api.elevenlabs.io/v1/text-to-speech/{self._voice_id}/stream"
 
         payload = {"text": text, "model_id": self._model}

src/pipecat/services/openai.py

@@ -23,6 +23,7 @@ from pipecat.frames.frames import (
     LLMFullResponseStartFrame,
     LLMMessagesFrame,
     LLMModelUpdateFrame,
+    MetricsFrame,
     TextFrame,
     URLImageRawFrame,
     VisionImageRawFrame
@@ -95,6 +96,7 @@ class BaseOpenAILLMService(LLMService):
             messages=messages,
             tools=context.tools,
             tool_choice=context.tool_choice,
+            stream_options={"include_usage": True}
         )
         return chunks
 
@@ -132,6 +134,19 @@ class BaseOpenAILLMService(LLMService):
         )
 
         async for chunk in chunk_stream:
+            if chunk.usage:
+                if self.can_generate_metrics() and self.metrics_enabled:
+                    tokens = {
+                        "processor": self.name,
+                        "prompt_tokens": chunk.usage.prompt_tokens,
+                        "completion_tokens": chunk.usage.completion_tokens,
+                        "total_tokens": chunk.usage.total_tokens
+                    }
+                    logger.debug(
+                        f"{self.name} prompt tokens: {tokens['prompt_tokens']}, completion tokens: {tokens['completion_tokens']}")
+
+                    await self.push_frame(MetricsFrame(tokens=[tokens]))
+
             if len(chunk.choices) == 0:
                 continue
 
@@ -323,7 +338,13 @@ class OpenAITTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
-
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
         try:
             await self.start_ttfb_metrics()
 

src/pipecat/services/playht.py

@@ -9,7 +9,7 @@ import struct
 
 from typing import AsyncGenerator
 
-from pipecat.frames.frames import AudioRawFrame, Frame
+from pipecat.frames.frames import AudioRawFrame, Frame, MetricsFrame
 from pipecat.services.ai_services import TTSService
 
 from loguru import logger
@@ -48,7 +48,13 @@ class PlayHTTTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
-
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
         try:
             b = bytearray()
             in_header = True

src/pipecat/services/xtts.py

@@ -8,7 +8,7 @@ import aiohttp
 
 from typing import Any, AsyncGenerator, Dict
 
-from pipecat.frames.frames import AudioRawFrame, ErrorFrame, Frame, StartFrame
+from pipecat.frames.frames import AudioRawFrame, ErrorFrame, Frame, MetricsFrame, StartFrame
 from pipecat.services.ai_services import TTSService
 
 from loguru import logger
@@ -70,7 +70,13 @@ class XTTSService(TTSService):
 
     async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
         logger.debug(f"Generating TTS: [{text}]")
-
+        if self.can_generate_metrics() and self.metrics_enabled:
+            characters = {
+                "processor": self.name,
+                "value": len(text),
+            }
+            logger.debug(f"{self.name} Characters: {characters['value']}")
+            await self.push_frame(MetricsFrame(characters=[characters]))
         if not self._studio_speakers:
             logger.error(f"{self} no studio speakers available")
             return

src/pipecat/transports/services/daily.py

@@ -699,6 +699,10 @@ class DailyOutputTransport(BaseOutputTransport):
             metrics["ttfb"] = frame.ttfb
         if frame.processing:
             metrics["processing"] = frame.processing
+        if frame.tokens:
+            metrics["tokens"] = frame.tokens
+        if frame.characters:
+            metrics["characters"] = frame.characters
 
         message = DailyTransportMessageFrame(message={
             "type": "pipecat-metrics",