Merge pull request #1717 from pipecat-ai/local_smart_turn_torch

Local smart turn torch

CHANGELOG.md

@@ -9,6 +9,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- Added support for cross-platform local smart turn detection. You can use 
+  `LocalSmartTurnAnalyzer` for on-device inference using Torch.
+
 - `BaseOutputTransport` now allows multiple destinations if the transport
   implementation supports it (e.g. Daily's custom tracks). With multiple
   destinations it is possible to send different audio or video tracks with a

examples/foundational/38b-smart-turn-local.py

@@ -0,0 +1,128 @@
+#
+# 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 pipecat.audio.turn.smart_turn.base_smart_turn import SmartTurnParams
+from pipecat.audio.turn.smart_turn.local_smart_turn import LocalSmartTurnAnalyzer
+from pipecat.audio.vad.silero import SileroVADAnalyzer
+from pipecat.audio.vad.vad_analyzer import VADParams
+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.openai.llm import OpenAILLMService
+from pipecat.transports.base_transport import TransportParams
+from pipecat.transports.network.small_webrtc import SmallWebRTCTransport
+from pipecat.transports.network.webrtc_connection import SmallWebRTCConnection
+
+load_dotenv(override=True)
+
+
+async def run_bot(webrtc_connection: SmallWebRTCConnection, _: argparse.Namespace):
+    logger.info(f"Starting bot")
+
+    # To use this locally, set the environment variable LOCAL_SMART_TURN_MODEL_PATH
+    # to the path where the smart-turn repo is cloned.
+    #
+    # Example setup:
+    #
+    #   # Git LFS (Large File Storage)
+    #   brew install git-lfs
+    #   # Hugging Face uses LFS to store large model files, including .mlpackage
+    #   git lfs install
+    #   # Clone the repo with the smart_turn_classifier.mlpackage
+    #   git clone https://huggingface.co/pipecat-ai/smart-turn
+    #
+    # Then set the env variable:
+    #   export LOCAL_SMART_TURN_MODEL_PATH=./smart-turn
+    # or add it to your .env file
+    smart_turn_model_path = os.getenv("LOCAL_SMART_TURN_MODEL_PATH")
+
+    transport = SmallWebRTCTransport(
+        webrtc_connection=webrtc_connection,
+        params=TransportParams(
+            audio_in_enabled=True,
+            audio_out_enabled=True,
+            vad_analyzer=SileroVADAnalyzer(params=VADParams(stop_secs=0.2)),
+            turn_analyzer=LocalSmartTurnAnalyzer(
+                smart_turn_model_path=smart_turn_model_path, params=SmartTurnParams()
+            ),
+        ),
+    )
+
+    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"))
+
+    messages = [
+        {
+            "role": "system",
+            "content": "You are a helpful LLM in a WebRTC call. Your goal is to demonstrate your capabilities in a succinct way. Your output will be converted to audio so don't include special characters in your answers. Respond to what the user said in a creative and helpful way.",
+        },
+    ]
+
+    context = OpenAILLMContext(messages)
+    context_aggregator = llm.create_context_aggregator(context)
+
+    pipeline = Pipeline(
+        [
+            transport.input(),  # Transport user input
+            stt,
+            context_aggregator.user(),  # User responses
+            llm,  # LLM
+            tts,  # TTS
+            transport.output(),  # Transport bot output
+            context_aggregator.assistant(),  # Assistant spoken responses
+        ]
+    )
+
+    task = PipelineTask(
+        pipeline,
+        params=PipelineParams(
+            allow_interruptions=True,
+            enable_metrics=True,
+            enable_usage_metrics=True,
+            report_only_initial_ttfb=True,
+        ),
+    )
+
+    @transport.event_handler("on_client_connected")
+    async def on_client_connected(transport, client):
+        logger.info(f"Client connected")
+        # Kick off the conversation.
+        messages.append({"role": "system", "content": "Please introduce yourself to the user."})
+        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")
+
+    @transport.event_handler("on_client_closed")
+    async def on_client_closed(transport, client):
+        logger.info(f"Client closed connection")
+        await task.cancel()
+
+    runner = PipelineRunner(handle_sigint=False)
+
+    await runner.run(task)
+
+
+if __name__ == "__main__":
+    from run import main
+
+    main()

src/pipecat/audio/turn/smart_turn/local_smart_turn.py

@@ -0,0 +1,73 @@
+#
+# Copyright (c) 2024–2025, Daily
+#
+# SPDX-License-Identifier: BSD 2-Clause License
+#
+
+
+from typing import Any, Dict
+
+import numpy as np
+from loguru import logger
+
+from pipecat.audio.turn.smart_turn.base_smart_turn import BaseSmartTurn
+
+try:
+    import torch
+    from transformers import AutoFeatureExtractor, Wav2Vec2BertForSequenceClassification
+except ModuleNotFoundError as e:
+    logger.error(f"Exception: {e}")
+    logger.error(
+        "In order to use the LocalSmartTurnAnalyzer, you need to `pip install pipecat-ai[local-smart-turn]`."
+    )
+    raise Exception(f"Missing module: {e}")
+
+
+class LocalSmartTurnAnalyzer(BaseSmartTurn):
+    def __init__(self, *, smart_turn_model_path: str, **kwargs):
+        super().__init__(**kwargs)
+
+        if not smart_turn_model_path:
+            # Define the path to the pretrained model on Hugging Face
+            smart_turn_model_path = "pipecat-ai/smart-turn"
+
+        logger.debug("Loading Local Smart Turn model...")
+        # Load the pretrained model for sequence classification
+        self._turn_model = Wav2Vec2BertForSequenceClassification.from_pretrained(
+            smart_turn_model_path
+        )
+        # Load the corresponding feature extractor for preprocessing audio
+        self._turn_processor = AutoFeatureExtractor.from_pretrained(smart_turn_model_path)
+        # Set device to GPU if available, else CPU
+        self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        # Move model to selected device and set it to evaluation mode
+        self._turn_model = self._turn_model.to(self._device)
+        self._turn_model.eval()
+        logger.debug("Loaded Local Smart Turn")
+
+    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+        inputs = self._turn_processor(
+            audio_array,
+            sampling_rate=16000,
+            padding="max_length",
+            truncation=True,
+            max_length=800,  # Maximum length as specified in training
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+
+        # Move input tensors to the same device as the model
+        inputs = {k: v.to(self._device) for k, v in inputs.items()}
+
+        # Disable gradient calculation for inference
+        with torch.no_grad():
+            outputs = self._turn_model(**inputs)
+            logits = outputs.logits
+            probabilities = torch.nn.functional.softmax(logits, dim=1)
+            completion_prob = probabilities[0, 1].item()  # Probability of class 1 (Complete)
+            prediction = 1 if completion_prob > 0.5 else 0
+
+        return {
+            "prediction": prediction,
+            "probability": completion_prob,
+        }