diff --git a/changelog/4244.fixed.md b/changelog/4244.fixed.md new file mode 100644 index 000000000..271e39c37 --- /dev/null +++ b/changelog/4244.fixed.md @@ -0,0 +1 @@ +- Fixed `VADController` getting stuck in the `SPEAKING` state when audio frames stop arriving mid-speech (e.g. user mutes mic). A new `audio_idle_timeout` parameter (default 1s, set to 0 to disable) forces a transition back to `QUIET` and emits `on_speech_stopped` when no audio is received while speaking. diff --git a/src/pipecat/audio/vad/vad_controller.py b/src/pipecat/audio/vad/vad_controller.py index 0316f3793..fefe3bec1 100644 --- a/src/pipecat/audio/vad/vad_controller.py +++ b/src/pipecat/audio/vad/vad_controller.py @@ -10,8 +10,11 @@ This module provides a controller that wraps a VADAnalyzer to track speech state and emit events when speech starts, stops, or is actively detected. """ +import asyncio import time -from typing import Type +from typing import Optional, Type + +from loguru import logger from pipecat.audio.vad.vad_analyzer import VADAnalyzer, VADState from pipecat.frames.frames import ( @@ -22,6 +25,7 @@ from pipecat.frames.frames import ( VADParamsUpdateFrame, ) from pipecat.processors.frame_processor import FrameDirection +from pipecat.utils.asyncio.task_manager import BaseTaskManager from pipecat.utils.base_object import BaseObject @@ -35,7 +39,8 @@ class VADController(BaseObject): Event handlers available: - on_speech_started: Called when speech begins. - - on_speech_stopped: Called when speech ends. + - on_speech_stopped: Called when speech ends, including forced stop when + the audio stream goes idle (no frames received while speaking). - on_speech_activity: Called periodically while speech is detected. - on_push_frame: Called when the controller wants to push a frame. - on_broadcast_frame: Called when the controller wants to broadcast a frame. @@ -63,30 +68,62 @@ class VADController(BaseObject): ... """ - def __init__(self, vad_analyzer: VADAnalyzer, *, speech_activity_period: float = 0.2): + def __init__( + self, + vad_analyzer: VADAnalyzer, + *, + speech_activity_period: float = 0.2, + audio_idle_timeout: float = 1.0, + ): """Initialize the VAD controller. Args: vad_analyzer: The `VADAnalyzer` instance for processing audio. speech_activity_period: Minimum interval in seconds between `on_speech_activity` events. Defaults to 0.2. + audio_idle_timeout: Timeout in seconds to force speech stop + when no audio frames are received while in SPEAKING state. + This handles cases like mic mute mid-speech. + Set to 0 to disable. Defaults to 1.0. """ super().__init__() self._vad_analyzer = vad_analyzer self._vad_state: VADState = VADState.QUIET + self._task_manager: Optional[BaseTaskManager] = None + # Last time a on_speech_activity was triggered. self._speech_activity_time = 0 # How often a on_speech_activity event should be triggered (value should # be greater than the audio chunks to have any effect). self._speech_activity_period = speech_activity_period + # Audio idle detection: force speech stop when no audio arrives + # while in SPEAKING state (e.g. user mutes mic mid-speech). + self._last_audio_time: float = 0.0 + self._audio_idle_timeout = audio_idle_timeout + self._audio_idle_task: Optional[asyncio.Task] = None + self._register_event_handler("on_speech_started", sync=True) self._register_event_handler("on_speech_stopped", sync=True) self._register_event_handler("on_speech_activity", sync=True) self._register_event_handler("on_push_frame", sync=True) self._register_event_handler("on_broadcast_frame", sync=True) + async def setup(self, task_manager: BaseTaskManager): + """Initialize the controller with the given task manager. + + Args: + task_manager: The task manager to be associated with this instance. + """ + self._task_manager = task_manager + self._last_audio_time = time.monotonic() + if self._audio_idle_timeout > 0 and not self._audio_idle_task: + self._audio_idle_task = self._task_manager.create_task( + self._audio_idle_handler(), + f"{self}::_audio_idle_handler", + ) + async def process_frame(self, frame: Frame): """Process a frame and handle VAD-related events. @@ -116,6 +153,10 @@ class VADController(BaseObject): It waits for all currently executing event handler tasks to finish before returning. """ + await super().cleanup() + if self._audio_idle_task and self._task_manager: + await self._task_manager.cancel_task(self._audio_idle_task) + self._audio_idle_task = None if self._vad_analyzer: await self._vad_analyzer.cleanup() @@ -128,6 +169,8 @@ class VADController(BaseObject): Args: frame: Audio frame to process. """ + self._last_audio_time = time.monotonic() + self._vad_state = await self._handle_vad(frame.audio, self._vad_state) if self._vad_state == VADState.SPEAKING: @@ -149,6 +192,29 @@ class VADController(BaseObject): vad_state = new_vad_state return vad_state + async def _audio_idle_handler(self): + """Monitor for an idle audio stream while in SPEAKING state. + + When no audio frames arrive for `audio_idle_timeout` seconds + (e.g. user mutes mic mid-speech), forces a transition to QUIET and + emits `on_speech_stopped`. + """ + while True: + deadline = self._last_audio_time + self._audio_idle_timeout + remaining = deadline - time.monotonic() + if remaining > 0: + # Audio is still recent; sleep only for the remaining window. + await asyncio.sleep(remaining) + continue + + if self._vad_state == VADState.SPEAKING: + logger.warning(f"{self}: no audio received while speaking, forcing speech stop") + self._vad_state = VADState.QUIET + await self._call_event_handler("on_speech_stopped") + + # Wait for the next potential idle window. + await asyncio.sleep(self._audio_idle_timeout) + async def _maybe_speech_activity(self): """Handle user speaking frame.""" diff_time = time.time() - self._speech_activity_time diff --git a/src/pipecat/processors/aggregators/llm_response_universal.py b/src/pipecat/processors/aggregators/llm_response_universal.py index 616636494..94e56a972 100644 --- a/src/pipecat/processors/aggregators/llm_response_universal.py +++ b/src/pipecat/processors/aggregators/llm_response_universal.py @@ -107,6 +107,9 @@ class LLMUserAggregatorParams: has been idle (not speaking) for this duration. Set to 0 to disable idle detection. vad_analyzer: Voice Activity Detection analyzer instance. + audio_idle_timeout: Timeout in seconds to force speech stop when + no audio frames are received while in SPEAKING state (e.g. user mutes + mic mid-speech). Set to 0 to disable. Defaults to 1.0. filter_incomplete_user_turns: Whether to filter out incomplete user turns. When enabled, the LLM outputs a turn completion marker at the start of each response: ✓ (complete), ○ (incomplete short), or ◐ (incomplete long). @@ -121,6 +124,7 @@ class LLMUserAggregatorParams: user_turn_stop_timeout: float = 5.0 user_idle_timeout: float = 0 vad_analyzer: Optional[VADAnalyzer] = None + audio_idle_timeout: float = 1.0 filter_incomplete_user_turns: bool = False user_turn_completion_config: Optional[UserTurnCompletionConfig] = None @@ -467,7 +471,10 @@ class LLMUserAggregator(LLMContextAggregator): # VAD controller self._vad_controller: Optional[VADController] = None if self._params.vad_analyzer: - self._vad_controller = VADController(self._params.vad_analyzer) + self._vad_controller = VADController( + self._params.vad_analyzer, + audio_idle_timeout=self._params.audio_idle_timeout, + ) self._vad_controller.add_event_handler("on_speech_started", self._on_vad_speech_started) self._vad_controller.add_event_handler("on_speech_stopped", self._on_vad_speech_stopped) self._vad_controller.add_event_handler( @@ -553,6 +560,9 @@ class LLMUserAggregator(LLMContextAggregator): return aggregation async def _start(self, frame: StartFrame): + if self._vad_controller: + await self._vad_controller.setup(self.task_manager) + await self._user_turn_controller.setup(self.task_manager) await self._user_idle_controller.setup(self.task_manager) @@ -584,6 +594,8 @@ class LLMUserAggregator(LLMContextAggregator): await self._cleanup() async def _cleanup(self): + if self._vad_controller: + await self._vad_controller.cleanup() await self._user_turn_controller.cleanup() await self._user_idle_controller.cleanup() diff --git a/src/pipecat/processors/audio/vad_processor.py b/src/pipecat/processors/audio/vad_processor.py index 9145f52cb..aaa769061 100644 --- a/src/pipecat/processors/audio/vad_processor.py +++ b/src/pipecat/processors/audio/vad_processor.py @@ -18,6 +18,7 @@ from pipecat.audio.vad.vad_analyzer import VADAnalyzer from pipecat.audio.vad.vad_controller import VADController from pipecat.frames.frames import ( Frame, + StartFrame, UserSpeakingFrame, VADUserStartedSpeakingFrame, VADUserStoppedSpeakingFrame, @@ -45,6 +46,7 @@ class VADProcessor(FrameProcessor): *, vad_analyzer: VADAnalyzer, speech_activity_period: float = 0.2, + audio_idle_timeout: float = 1.0, **kwargs, ): """Initialize the VAD processor. @@ -53,11 +55,16 @@ class VADProcessor(FrameProcessor): vad_analyzer: The VADAnalyzer instance for processing audio. speech_activity_period: Minimum interval in seconds between UserSpeakingFrame pushes. Defaults to 0.2. + audio_idle_timeout: Timeout in seconds to force speech stop + when no audio frames are received while in SPEAKING state. + Set to 0 to disable. Defaults to 1.0. **kwargs: Additional arguments passed to parent class. """ super().__init__(**kwargs) self._vad_controller = VADController( - vad_analyzer, speech_activity_period=speech_activity_period + vad_analyzer, + speech_activity_period=speech_activity_period, + audio_idle_timeout=audio_idle_timeout, ) # Push VAD frames when speech events are detected @@ -90,6 +97,11 @@ class VADProcessor(FrameProcessor): async def on_broadcast_frame(_controller, frame_cls: Type[Frame], **kwargs): await self.broadcast_frame(frame_cls, **kwargs) + async def cleanup(self): + """Clean up VAD controller resources.""" + await super().cleanup() + await self._vad_controller.cleanup() + async def process_frame(self, frame: Frame, direction: FrameDirection): """Process a frame through VAD and forward it. @@ -104,5 +116,8 @@ class VADProcessor(FrameProcessor): # 2. Audio flows through immediately while VAD detection happens after await self.push_frame(frame, direction) + if isinstance(frame, StartFrame): + await self._vad_controller.setup(self.task_manager) + # Let the VAD controller handle the frame await self._vad_controller.process_frame(frame) diff --git a/tests/test_vad_controller.py b/tests/test_vad_controller.py index c208b5122..30f2fe56c 100644 --- a/tests/test_vad_controller.py +++ b/tests/test_vad_controller.py @@ -4,6 +4,7 @@ # SPDX-License-Identifier: BSD 2-Clause License # +import asyncio import unittest from typing import List @@ -11,6 +12,7 @@ from pipecat.audio.vad.vad_analyzer import VADAnalyzer, VADParams, VADState from pipecat.audio.vad.vad_controller import VADController from pipecat.frames.frames import Frame, InputAudioRawFrame, SpeechControlParamsFrame, StartFrame from pipecat.processors.frame_processor import FrameDirection +from pipecat.utils.asyncio.task_manager import TaskManager, TaskManagerParams class MockVADAnalyzer(VADAnalyzer): @@ -206,5 +208,64 @@ class TestVADController(unittest.IsolatedAsyncioTestCase): self.assertIsInstance(broadcast_calls[0][1]["vad_params"], VADParams) +AUDIO_IDLE_TIMEOUT = 0.1 + + +class TestVADControllerAudioIdle(unittest.IsolatedAsyncioTestCase): + async def asyncSetUp(self): + self.task_manager = TaskManager() + self.task_manager.setup(TaskManagerParams(loop=asyncio.get_running_loop())) + + async def test_audio_idle_forces_speech_stop(self): + """Test that on_speech_stopped fires when no audio arrives while SPEAKING.""" + analyzer = MockVADAnalyzer() + controller = VADController(analyzer, audio_idle_timeout=AUDIO_IDLE_TIMEOUT) + + speech_stopped = False + + @controller.event_handler("on_speech_stopped") + async def on_speech_stopped(_controller): + nonlocal speech_stopped + speech_stopped = True + + start_frame = StartFrame(audio_in_sample_rate=16000, audio_out_sample_rate=16000) + await controller.process_frame(start_frame) + await controller.setup(self.task_manager) + + # Enter SPEAKING state + audio_frame = InputAudioRawFrame(audio=b"\x00" * 1024, sample_rate=16000, num_channels=1) + analyzer.set_next_state(VADState.SPEAKING) + await controller.process_frame(audio_frame) + self.assertFalse(speech_stopped) + + # Stop sending audio, wait for idle timeout + await asyncio.sleep(AUDIO_IDLE_TIMEOUT + 0.1) + self.assertTrue(speech_stopped) + + await controller.cleanup() + + async def test_audio_idle_does_not_fire_when_quiet(self): + """Test that idle timeout does NOT fire when VAD is in QUIET state.""" + analyzer = MockVADAnalyzer() + controller = VADController(analyzer, audio_idle_timeout=AUDIO_IDLE_TIMEOUT) + + speech_stopped = False + + @controller.event_handler("on_speech_stopped") + async def on_speech_stopped(_controller): + nonlocal speech_stopped + speech_stopped = True + + start_frame = StartFrame(audio_in_sample_rate=16000, audio_out_sample_rate=16000) + await controller.process_frame(start_frame) + await controller.setup(self.task_manager) + + # Stay in QUIET state, wait past idle timeout + await asyncio.sleep(AUDIO_IDLE_TIMEOUT + 0.1) + self.assertFalse(speech_stopped) + + await controller.cleanup() + + if __name__ == "__main__": unittest.main()