BaseSmartTurn: process speech in a separate thread

CHANGELOG.md

@@ -11,6 +11,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Include OpenAI-based LLM services cached tokens to `MetricsFrame`.
 
+## Fixed
+
+- Fixed an issue where local SmartTurn was not being ran in a separate thread.
+
 ## [0.0.86] - 2025-09-24
 
 ### Added

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

@@ -11,8 +11,10 @@ machine learning models to determine when a user has finished speaking, going
 beyond simple silence-based detection.
 """
 
+import asyncio
 import time
 from abc import abstractmethod
+from concurrent.futures import ThreadPoolExecutor
 from typing import Any, Dict, Optional, Tuple
 
 import numpy as np
@@ -77,6 +79,9 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         self._speech_triggered = False
         self._silence_ms = 0
         self._speech_start_time = 0
+        # Thread executor that will run the model. We only need one thread per
+        # analyzer because one analyzer just handles one audio stream.
+        self._executor = ThreadPoolExecutor(max_workers=1)
 
     @property
     def speech_triggered(self) -> bool:
@@ -151,7 +156,10 @@ class BaseSmartTurn(BaseTurnAnalyzer):
             Tuple containing the end-of-turn state and optional metrics data
             from the ML model analysis.
         """
-        state, result = await self._process_speech_segment(self._audio_buffer)
+        loop = asyncio.get_running_loop()
+        state, result = await loop.run_in_executor(
+            self._executor, self._process_speech_segment, self._audio_buffer
+        )
         if state == EndOfTurnState.COMPLETE or USE_ONLY_LAST_VAD_SEGMENT:
             self._clear(state)
         logger.debug(f"End of Turn result: {state}")
@@ -169,9 +177,7 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         self._speech_start_time = 0
         self._silence_ms = 0
 
-    async def _process_speech_segment(
-        self, audio_buffer
-    ) -> Tuple[EndOfTurnState, Optional[MetricsData]]:
+    def _process_speech_segment(self, audio_buffer) -> Tuple[EndOfTurnState, Optional[MetricsData]]:
         """Process accumulated audio segment using ML model."""
         state = EndOfTurnState.INCOMPLETE
 
@@ -203,7 +209,7 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         if len(segment_audio) > 0:
             start_time = time.perf_counter()
             try:
-                result = await self._predict_endpoint(segment_audio)
+                result = self._predict_endpoint(segment_audio)
                 state = (
                     EndOfTurnState.COMPLETE
                     if result["prediction"] == 1
@@ -249,6 +255,6 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         return state, result_data
 
     @abstractmethod
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using ML model from audio data."""
         pass

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

@@ -104,11 +104,15 @@ class HttpSmartTurnAnalyzer(BaseSmartTurn):
             logger.error(f"Failed to send raw request to Daily Smart Turn: {e}")
             raise Exception("Failed to send raw request to Daily Smart Turn.")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using remote HTTP ML service."""
         try:
             serialized_array = self._serialize_array(audio_array)
-            return await self._send_raw_request(serialized_array)
+            loop = asyncio.get_running_loop()
+            future = asyncio.run_coroutine_threadsafe(
+                self._send_raw_request(serialized_array), loop
+            )
+            return future.result()
         except Exception as e:
             logger.error(f"Smart turn prediction failed: {str(e)}")
             # Return an incomplete prediction when a failure occurs

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

@@ -64,7 +64,7 @@ class LocalSmartTurnAnalyzer(BaseSmartTurn):
         self._turn_model.eval()
         logger.debug("Loaded Local Smart Turn")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using local PyTorch model."""
         inputs = self._turn_processor(
             audio_array,

src/pipecat/audio/turn/smart_turn/local_smart_turn_v2.py

@@ -73,7 +73,7 @@ class LocalSmartTurnAnalyzerV2(BaseSmartTurn):
         self._turn_model.eval()
         logger.debug("Loaded Local Smart Turn v2")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using local PyTorch model."""
         inputs = self._turn_processor(
             audio_array,

src/pipecat/audio/turn/smart_turn/local_smart_turn_v3.py

@@ -77,7 +77,7 @@ class LocalSmartTurnAnalyzerV3(BaseSmartTurn):
 
         logger.debug("Loaded Local Smart Turn v3")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using local ONNX model."""
 
         def truncate_audio_to_last_n_seconds(audio_array, n_seconds=8, sample_rate=16000):