feat: integra VAD ao controle de turnos da conversa

2026-11-03 17:44:30 -03:00
parent d27c26c166
commit bdd2f6f67d
9 changed files with 770 additions and 24 deletions
--- a/README.md
+++ b/README.md
@@ -388,7 +388,7 @@ The script reads `audios/chafik_normalize.wav` (16 kHz, mono, 16-bit PCM), sends
 config/                         Environment YAML and provider catalog
 src/voice_assistant/
  audio_gateway/app/            WebSocket endpoint and protocol
-  application/                  VoiceSession and transport-independent events
+  application/                  VoiceSession, continuous PCM history, and application events
  domain/                       Session state and domain rules
  vad/                          VAD port, speech boundaries, and local Silero adapter
  audio/                        Canonical audio models
@@ -418,7 +418,7 @@ tools/                          Integration and smoke scripts



-## Backend VAD contract (integration pending)
+## Backend VAD contract

 `vad/adapter.py` defines one stateful detector per continuous input stream.
 Its input is the canonical PCM S16LE, mono, 16 kHz audio format. Arbitrary
@@ -432,9 +432,9 @@ boundary. Model padding never extends the input timeline.
 `reset()` discards detection state and restarts the timeline, and `aclose()`
 releases resources without emitting new speech events. Calls are serialized.
 These contracts do not start turns, select providers, or carry transport IDs.
-The local Silero implementation is available independently of the running
-conversation. Application integration and browser VAD replacement are subsequent
-tasks; the active conversation flow still uses its existing browser VAD.
+The Silero adapter can control turns through the continuous VoiceSession API.
+Browser/WebSocket migration is the next task; the existing web client continues
+to use its current turn commands until that migration.


 ### Local Silero VAD adapter
@@ -497,5 +497,49 @@ remains compatible with existing configurations; explicitly creating VAD in
 such an environment raises a configuration error. Changing YAML values affects
 new adapter instances after configuration is reloaded; there is no live reload.

-This factory makes the configuration usable independently. Connecting it to the
-conversation lifecycle remains the next task.
+The VAD factory can be supplied to continuous voice sessions. The existing web
+client remains on manual turn commands until the browser migration.
+
+
+### Continuous audio and backend turn control
+
+A `VoiceSession` constructed with a VAD receives continuous canonical PCM through
+`await session.receive_audio(pcm)`. The caller supplies audio without a turn ID;
+the channel owns transport sequencing. `SpeechStarted` opens an STT turn and
+`SpeechEnded` schedules normal STT finalization in the background. New speech
+interrupts an active response/finalization before opening a fresh turn. Empty
+STT results release the turn through an interruption event rather than leaving
+an active turn behind. The VAD is not reset between conversation turns.
+
+The application's sample-indexed audio history preserves the beginning of speech
+when VAD confirmation arrives later. PCM already sent to a turn is not replayed
+within that turn; pre-roll never includes the previous utterance. History and
+input size are bounded by environment configuration:
+
+```yaml
+voice_input:
+  pre_roll_ms: 200
+  history_ms: 2000
+  max_input_chunk_bytes: 65536
+```
+
+History is trimmed after processing each input chunk; transient storage is at most
+retained history plus one accepted chunk. A speech start older than retained
+history is reported as an input error instead of silently dropping speech.
+Invalid PCM is rejected before detection or timeline advancement. A processing
+failure requires a new session; cleanup still releases VAD and provider resources.
+
+Use `finish_audio_input()` for normal end of a continuous stream. Session end or
+disconnect aborts through `end()`/`aclose()`, without forcing a final speech event.
+The caller shields cleanup from transport cancellation, as for manual sessions.
+
+Backend-controlled sessions emit `TurnOpened` and `TurnInterrupted`; the WebSocket
+mapper translates them to the existing `turn.started` and `interrupted` events.
+Adapters should discard outdated queued playback when a new turn opens: backend
+turn completion describes produced output, not physical speaker completion.
+
+Manual `start_turn`/`send_audio`/`stop_turn` commands remain available on sessions
+without VAD. Mixing those commands with a VAD-controlled session is rejected.
+`create_voice_session(..., server_vad=True)` composes the continuous backend API;
+the current `/ws/audio` receive loop still uses the manual mode. The next task
+connects continuous input to that route and simplifies the browser accordingly.
--- a/config/development.yaml
+++ b/config/development.yaml
@@ -24,3 +24,8 @@ vad:
    min_silence_duration_ms: 300
    speech_pad_ms: 30
    worker_timeout_seconds: 10.0
+
+voice_input:
+  pre_roll_ms: 200
+  history_ms: 2000
+  max_input_chunk_bytes: 65536
--- a/src/voice_assistant/application/audio_history.py
+++ b/src/voice_assistant/application/audio_history.py
@@ -0,0 +1,32 @@
+"""A bounded PCM history addressed by absolute input sample positions."""
+
+from voice_assistant.audio.models import CANONICAL_AUDIO_FORMAT
+
+
+class AudioHistory:
+    def __init__(self, history_ms: int) -> None:
+        if type(history_ms) is not int or history_ms <= 0:
+            raise ValueError("history_ms must be a positive integer.")
+        self._retained_samples = history_ms * CANONICAL_AUDIO_FORMAT.sample_rate // 1000
+        self._audio = bytearray()
+        self.start_sample = 0
+        self.end_sample = 0
+
+    def append(self, audio: bytes) -> None:
+        self._audio.extend(audio)
+        self.end_sample += len(audio) // CANONICAL_AUDIO_FORMAT.frame_size_bytes
+
+    def read(self, start: int, end: int) -> bytes:
+        if not self.start_sample <= start <= end <= self.end_sample:
+            raise ValueError("Requested audio is outside the retained input history.")
+        width = CANONICAL_AUDIO_FORMAT.frame_size_bytes
+        return bytes(self._audio[(start - self.start_sample) * width:(end - self.start_sample) * width])
+
+    def trim(self) -> None:
+        start = max(self.start_sample, self.end_sample - self._retained_samples)
+        del self._audio[:(start - self.start_sample) * CANONICAL_AUDIO_FORMAT.frame_size_bytes]
+        self.start_sample = start
+
+    def clear(self) -> None:
+        self._audio.clear()
+        self.start_sample = self.end_sample
--- a/src/voice_assistant/application/events.py
+++ b/src/voice_assistant/application/events.py
@@ -41,6 +41,18 @@ class SpeechChunk:
    audio_format: AudioFormat


+@dataclass(frozen=True, slots=True)
+class TurnOpened:
+    session_id: str
+    turn_id: int
+
+
+@dataclass(frozen=True, slots=True)
+class TurnInterrupted:
+    session_id: str
+    turn_id: int
+
+
@dataclass(frozen=True, slots=True)
 class TurnCompleted:
    session_id: str
@@ -58,5 +70,5 @@ class SessionFailure:

 type VoiceEvent = (
    TranscriptPartial | TranscriptFinal | AgentTextDelta | AgentReply
-    | SpeechChunk | TurnCompleted | SessionFailure
+    | SpeechChunk | TurnOpened | TurnInterrupted | TurnCompleted | SessionFailure
 )
--- a/src/voice_assistant/application/voice_session.py
+++ b/src/voice_assistant/application/voice_session.py
@@ -6,7 +6,12 @@ from collections.abc import Awaitable, Callable
 from contextlib import aclosing

 from voice_assistant.agent.adapter import StreamingAgentAdapter
+from voice_assistant.application.audio_history import AudioHistory
+from voice_assistant.vad.adapter import VADAdapter
+from voice_assistant.vad.models import SpeechStarted, SpeechEnded, VADEvent
 from voice_assistant.application.events import (
+    TurnOpened,
+    TurnInterrupted,
    AgentReply,
    AgentTextDelta,
    SessionFailure,
@@ -18,6 +23,7 @@ from voice_assistant.application.events import (
 )
 from voice_assistant.audio.models import AudioFormat, CANONICAL_AUDIO_FORMAT
 from voice_assistant.stt.adapter import STTAdapter
+from voice_assistant.turn_manager import TurnState
 from voice_assistant.voice_pipeline import (
    PipelineAgentResponse,
    PipelineAudioChunk,
@@ -64,9 +70,30 @@ class VoiceSession:
        emit: EventSink,
        stt_stop_timeout: float = 10.0,
        finalization_cancel_timeout: float = 5.0,
+        vad: VADAdapter | None = None,
+        pre_roll_ms: int = 200,
+        history_ms: int = 2000,
+        max_input_chunk_bytes: int = 65536,
    ) -> None:
        if stt_stop_timeout <= 0 or finalization_cancel_timeout <= 0:
            raise ValueError("Session timeouts must be positive.")
+        if type(pre_roll_ms) is not int or pre_roll_ms < 0:
+            raise ValueError("pre_roll_ms must be a nonnegative integer.")
+        if type(history_ms) is not int or history_ms <= 0 or pre_roll_ms > history_ms:
+            raise ValueError("history_ms must be positive and cover pre_roll_ms.")
+        if type(max_input_chunk_bytes) is not int or max_input_chunk_bytes <= 0:
+            raise ValueError("max_input_chunk_bytes must be a positive integer.")
+        self._vad = vad
+        self._input_lock = asyncio.Lock()
+        self._history = AudioHistory(history_ms)
+        self._pre_roll_samples = pre_roll_ms * CANONICAL_AUDIO_FORMAT.sample_rate // 1000
+        self._max_input_chunk_bytes = max_input_chunk_bytes
+        self._input_failed = False
+        self._input_finished = False
+        self._closing = False
+        self._capture_turn_id: int | None = None
+        self._sent_until = 0
+        self._last_voice_end = 0
        self._pipeline = pipeline
        self._create_stt = create_stt
        self._emit = emit
@@ -116,7 +143,17 @@ class VoiceSession:
                "SESSION_ALREADY_CLOSED", "This session is already closed."
            )

+    def _require_manual_input(self) -> None:
+        if self._vad is not None:
+            raise VoiceSessionError(
+                "TURN_CONTROLLED_BY_VAD", "This session uses server-side turn detection.",
+            )
+
    async def start_turn(self) -> int:
+        self._require_manual_input()
+        return await self._start_turn()
+
+    async def _start_turn(self) -> int:
        self._require_open()
        if self.active_turn_id is not None:
            raise VoiceSessionError(
@@ -168,6 +205,12 @@ class VoiceSession:

    async def send_audio(
        self, turn_id: int, sequence_number: int, audio: bytes,
+    ) -> None:
+        self._require_manual_input()
+        await self._send_audio(turn_id, sequence_number, audio)
+
+    async def _send_audio(
+        self, turn_id: int, sequence_number: int, audio: bytes,
    ) -> None:
        self.validate_audio_input(turn_id)
        CANONICAL_AUDIO_FORMAT.validate_chunk(audio)
@@ -187,6 +230,10 @@ class VoiceSession:
            ) from exc

    def stop_turn(self, turn_id: int) -> None:
+        self._require_manual_input()
+        self._stop_turn(turn_id)
+
+    def _stop_turn(self, turn_id: int) -> None:
        self._require_open()
        if self._stt is None:
            raise RuntimeError("STT session has not been started.")
@@ -200,6 +247,12 @@ class VoiceSession:
        )

    async def interrupt(self, turn_id: int) -> None:
+        async with self._input_lock:
+            await self._interrupt(turn_id)
+            if self._capture_turn_id == turn_id:
+                self._capture_turn_id = None
+
+    async def _interrupt(self, turn_id: int) -> None:
        self._require_open()
        if not self._pipeline.is_active_turn(turn_id):
            raise VoiceSessionError(
@@ -221,6 +274,11 @@ class VoiceSession:
            ) from exc

    async def end(self) -> None:
+        self._closing = True
+        async with self._input_lock:
+            await self._end()
+
+    async def _end(self) -> None:
        self._require_open()
        try:
            async with self._turn_transition_lock:
@@ -306,6 +364,12 @@ class VoiceSession:
                    or not self._pipeline.is_active_turn(turn_id)
                ):
                    return
+                turn = self._pipeline.turn_manager.active_turn
+                if self._vad is not None and turn is not None and turn.state == TurnState.LISTENING:
+                    # A VAD trigger may produce no recognizable transcript.
+                    self._pipeline.interrupt_if_active(turn_id)
+                    await self._emit(TurnInterrupted(self.session_id, turn_id))
+                    return
                self._pipeline.complete_turn(turn_id)
                self._completed_turns.add(turn_id)
                await self._emit(TurnCompleted(self.session_id, turn_id))
@@ -389,6 +453,11 @@ class VoiceSession:
                self._stt_events_task = None

    async def aclose(self) -> None:
+        self._closing = True
+        async with self._input_lock:
+            await self._close_resources()
+
+    async def _close_resources(self) -> None:
        """Attempt every cleanup operation, including after partial startup.

        The caller must shield this awaited cleanup from transport-scope
@@ -425,3 +494,113 @@ class VoiceSession:
                    "Failed to close %s provider. session_id=%s",
                    name, self.session_id,
                )
+
+        if self._vad is not None:
+            try:
+                await self._vad.aclose()
+            except Exception:
+                logger.exception("Failed to close VAD. session_id=%s", self.session_id)
+        self._history.clear()
+        self._capture_turn_id = None
+
+    def _require_continuous_input(self) -> VADAdapter:
+        self._require_open()
+        if self._vad is None:
+            raise VoiceSessionError("INVALID_INPUT_MODE", "Continuous input requires VAD.")
+        if self._closing or self._input_finished:
+            raise VoiceSessionError("INPUT_CLOSED", "Continuous audio input is closed.")
+        if self._input_failed:
+            raise VoiceSessionError("VOICE_INPUT_FAILED", "Create a new session after input failure.")
+        return self._vad
+
+    async def receive_audio(self, audio: bytes) -> None:
+        """Receive continuous PCM independently of client-managed turn IDs.
+
+        The channel serializes and validates its own packet sequence. This
+        operation validates PCM before advancing the application's timeline.
+        """
+        CANONICAL_AUDIO_FORMAT.validate_chunk(audio)
+        if len(audio) > self._max_input_chunk_bytes:
+            raise ValueError("Continuous audio chunk exceeds its configured size limit.")
+        async with self._input_lock:
+            vad = self._require_continuous_input()
+            self._history.append(audio)
+            try:
+                events = await vad.process(audio)
+                if self._closing:
+                    raise VoiceSessionError("INPUT_CLOSED", "Session shutdown is in progress.")
+                await self._handle_voice_activity(events)
+                await self._forward_input_until(self._history.end_sample)
+            except asyncio.CancelledError:
+                self._input_failed = True
+                raise
+            except Exception as exc:
+                self._input_failed = True
+                if isinstance(exc, VoiceSessionError):
+                    raise
+                raise VoiceSessionError(
+                    "VOICE_INPUT_FAILED", str(exc), turn_id=self.active_turn_id,
+                ) from exc
+            finally:
+                self._history.trim()
+
+    async def finish_audio_input(self) -> None:
+        """Finish a normal input stream; session end/disconnect aborts instead."""
+        async with self._input_lock:
+            vad = self._require_continuous_input()
+            try:
+                events = await vad.flush()
+                if self._closing:
+                    raise VoiceSessionError("INPUT_CLOSED", "Session shutdown is in progress.")
+                await self._handle_voice_activity(events)
+                self._input_finished = True
+            except asyncio.CancelledError:
+                self._input_failed = True
+                raise
+            except Exception as exc:
+                self._input_failed = True
+                if isinstance(exc, VoiceSessionError):
+                    raise
+                raise VoiceSessionError(
+                    "VOICE_INPUT_FAILED", str(exc), turn_id=self.active_turn_id,
+                ) from exc
+            finally:
+                self._history.trim()
+
+    async def _handle_voice_activity(self, events: tuple[VADEvent, ...]) -> None:
+        for event in events:
+            if isinstance(event, SpeechStarted):
+                if self._capture_turn_id is not None:
+                    raise ValueError("VAD emitted a duplicate speech start.")
+                if not self._history.start_sample <= event.sample_offset < self._history.end_sample:
+                    raise ValueError("Speech start is outside retained audio; increase history_ms.")
+                previous = self.active_turn_id
+                if previous is not None:
+                    await self._interrupt(previous)
+                    await self._emit(TurnInterrupted(self.session_id, previous))
+                else:
+                    await self._cancel_finalization_task()
+                turn_id = await self._start_turn()
+                self._capture_turn_id = turn_id
+                self._sent_until = max(
+                    self._history.start_sample, self._last_voice_end,
+                    event.sample_offset - self._pre_roll_samples,
+                )
+                await self._emit(TurnOpened(self.session_id, turn_id))
+            elif isinstance(event, SpeechEnded):
+                if not 0 <= event.sample_offset <= self._history.end_sample:
+                    raise ValueError("Speech end is outside input audio.")
+                self._last_voice_end = event.sample_offset
+                if self._capture_turn_id is not None:
+                    await self._forward_input_until(event.sample_offset)
+                    turn_id = self._capture_turn_id
+                    self._capture_turn_id = None
+                    self._stop_turn(turn_id)
+
+    async def _forward_input_until(self, end_sample: int) -> None:
+        turn_id = self._capture_turn_id
+        if turn_id is None or end_sample <= self._sent_until:
+            return
+        audio = self._history.read(self._sent_until, end_sample)
+        await self._send_audio(turn_id, self.expected_audio_sequence, audio)
+        self._sent_until = end_sample
--- a/src/voice_assistant/audio_gateway/app/main.py
+++ b/src/voice_assistant/audio_gateway/app/main.py
@@ -13,7 +13,7 @@ from pydantic import BaseModel, ValidationError

 from voice_assistant.application.events import (
    AgentReply, AgentTextDelta, SessionFailure, SpeechChunk,
-    TranscriptFinal, TranscriptPartial, TurnCompleted, VoiceEvent,
+    TranscriptFinal, TranscriptPartial, TurnCompleted, TurnOpened, TurnInterrupted, VoiceEvent,
 )
 from voice_assistant.application.voice_session import VoiceSession, VoiceSessionError
 from voice_assistant.audio.models import CANONICAL_AUDIO_FORMAT
@@ -26,9 +26,8 @@ from voice_assistant.audio_gateway.app.protocol import (
 from voice_assistant.config.loader import load_config, load_provider_catalog
 from voice_assistant.domain.session import SessionState
 from voice_assistant.providers import (
-    create_agent_adapter, create_stt_adapter, create_tts_adapter,
+    create_agent_adapter, create_stt_adapter, create_tts_adapter, create_vad_adapter,
 )
-from voice_assistant.tts.adapter import TTSAdapter
 from voice_assistant.turn_manager import TurnManager
 from voice_assistant.voice_pipeline import VoicePipeline

@@ -81,6 +80,10 @@ async def send_voice_event(websocket: WebSocket, event: VoiceEvent) -> None:
            payload = AgentDelta(session_id=sid, turn_id=tid, text=text)
        case AgentReply(session_id=sid, turn_id=tid, text=text):
            payload = AgentResponse(session_id=sid, turn_id=tid, text=text)
+        case TurnOpened(session_id=sid, turn_id=tid):
+            payload = TurnStarted(session_id=sid, turn_id=tid)
+        case TurnInterrupted(session_id=sid, turn_id=tid):
+            payload = Interrupted(session_id=sid, turn_id=tid)
        case TurnCompleted(session_id=sid, turn_id=tid):
            payload = TurnEnded(session_id=sid, turn_id=tid)
        case SessionFailure():
@@ -93,20 +96,34 @@ async def send_voice_event(websocket: WebSocket, event: VoiceEvent) -> None:
    await send_json(websocket, payload)


-async def create_voice_session(websocket: WebSocket, session_id: str | None) -> VoiceSession:
-    """Compose providers at the channel boundary, injecting application ports."""
+async def create_voice_session(
+    websocket: WebSocket, session_id: str | None, *, server_vad: bool = False,
+) -> VoiceSession:
+    """Compose manual input or continuous input controlled by backend VAD."""
    state = SessionState(session_id=session_id) if session_id is not None else SessionState()
-    agent = create_agent_adapter(config=app_config, catalog=provider_catalog)
-    tts: TTSAdapter | None = None
+    agent = None
+    tts = None
+    vad = None
    try:
+        agent = create_agent_adapter(config=app_config, catalog=provider_catalog)
        tts = create_tts_adapter(config=app_config, catalog=provider_catalog)
+        if server_vad:
+            vad = create_vad_adapter(config=app_config, catalog=provider_catalog)
        pipeline = VoicePipeline(
            session=state, turn_manager=create_turn_manager(), agent=agent, tts=tts,
        )
+        return VoiceSession(
+            pipeline=pipeline,
+            create_stt=partial(create_stt_adapter, config=app_config, catalog=provider_catalog),
+            emit=partial(send_voice_event, websocket),
+            stt_stop_timeout=STT_STOP_TIMEOUT_SECONDS,
+            finalization_cancel_timeout=FINALIZATION_CANCEL_TIMEOUT_SECONDS,
+            vad=vad,
+            **app_config.voice_input,
+        )
    except BaseException:
-        # A partially composed application has not yet acquired ownership.
        with anyio.CancelScope(shield=True):
-            for name, provider in (("Agent", agent), ("TTS", tts)):
+            for name, provider in (("Agent", agent), ("TTS", tts), ("VAD", vad)):
                if provider is None:
                    continue
                try:
@@ -116,13 +133,6 @@ async def create_voice_session(websocket: WebSocket, session_id: str | None) ->
                        "Failed to close %s after session composition failure.", name,
                    )
        raise
-    return VoiceSession(
-        pipeline=pipeline,
-        create_stt=partial(create_stt_adapter, config=app_config, catalog=provider_catalog),
-        emit=partial(send_voice_event, websocket),
-        stt_stop_timeout=STT_STOP_TIMEOUT_SECONDS,
-        finalization_cancel_timeout=FINALIZATION_CANCEL_TIMEOUT_SECONDS,
-    )


@app.websocket("/ws/audio")
--- a/src/voice_assistant/config/models.py
+++ b/src/voice_assistant/config/models.py
@@ -72,3 +72,4 @@ class AppConfig(BaseModel):
    tts: ComponentConfig
    agent: ComponentConfig
    vad: ComponentConfig | None = None
+    voice_input: dict[str, Any] = Field(default_factory=dict)
--- a/tests/test_vad_session_composition.py
+++ b/tests/test_vad_session_composition.py
@@ -0,0 +1,93 @@
+"""Composition selects VAD input and routes application events to the channel."""
+
+import json
+from pathlib import Path
+from unittest.mock import AsyncMock
+
+import pytest
+
+from voice_assistant.agent.mock import MockAgentAdapter
+from voice_assistant.application.events import TurnOpened, TurnInterrupted
+from voice_assistant.audio_gateway.app import main
+from voice_assistant.config.loader import load_config
+from voice_assistant.tts.mock import MockTTSAdapter
+from voice_assistant.vad.adapter import VADAdapter
+
+
+@pytest.fixture
+def anyio_backend():
+    return 'asyncio'
+
+
+class VAD(VADAdapter):
+    async def process(self, audio):
+        return ()
+
+    async def flush(self):
+        return ()
+
+    async def reset(self):
+        pass
+
+    async def aclose(self):
+        pass
+
+
+@pytest.mark.anyio
+@pytest.mark.parametrize('server_vad', [False, True])
+async def test_composition_selects_vad_and_accepts_environment_policy(monkeypatch, server_vad):
+    config = load_config('development', config_dir=Path(__file__).resolve().parents[1] / 'config')
+    assert config.voice_input == {
+        'pre_roll_ms': 200, 'history_ms': 2000, 'max_input_chunk_bytes': 65536,
+    }
+    monkeypatch.setattr(main, 'app_config', config)
+    monkeypatch.setattr(main, 'create_agent_adapter', lambda **kwargs: MockAgentAdapter())
+    monkeypatch.setattr(main, 'create_tts_adapter', lambda **kwargs: MockTTSAdapter())
+    detector = VAD()
+    calls = []
+
+    def factory(**kwargs):
+        calls.append(kwargs)
+        return detector
+
+    monkeypatch.setattr(main, 'create_vad_adapter', factory)
+    session = await main.create_voice_session(AsyncMock(), None, server_vad=server_vad)
+    try:
+        if server_vad:
+            assert len(calls) == 1
+            assert calls[0]['config'] is config
+            await session.receive_audio(bytes(512))
+        else:
+            assert calls == []
+            with pytest.raises(Exception, match='requires VAD'):
+                await session.receive_audio(bytes(512))
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_bad_policy_releases_all_created_providers(monkeypatch):
+    config = main.app_config.model_copy(update={'voice_input': {'history_ms': 0}})
+    agent, tts, vad = MockAgentAdapter(), MockTTSAdapter(), VAD()
+    for provider in [agent, tts, vad]:
+        provider.aclose = AsyncMock()
+    monkeypatch.setattr(main, 'app_config', config)
+    monkeypatch.setattr(main, 'create_agent_adapter', lambda **kwargs: agent)
+    monkeypatch.setattr(main, 'create_tts_adapter', lambda **kwargs: tts)
+    monkeypatch.setattr(main, 'create_vad_adapter', lambda **kwargs: vad)
+    with pytest.raises(ValueError, match='history_ms'):
+        await main.create_voice_session(AsyncMock(), None, server_vad=True)
+    for provider in [agent, tts, vad]:
+        provider.aclose.assert_awaited_once()
+
+
+@pytest.mark.anyio
+async def test_automatic_turn_events_map_to_existing_protocol():
+    socket = AsyncMock()
+    await main.send_voice_event(socket, TurnOpened('session', 2))
+    await main.send_voice_event(socket, TurnInterrupted('session', 1))
+    events = [json.loads(call.args[0]) for call in socket.send_text.await_args_list]
+    assert events == [
+        {'type': 'turn.started', 'session_id': 'session', 'turn_id': 2},
+        {'type': 'interrupted', 'session_id': 'session', 'turn_id': 1},
+    ]
--- a/tests/test_voice_session_vad.py
+++ b/tests/test_voice_session_vad.py
@@ -0,0 +1,370 @@
+"""Continuous, channel-independent PCM controls conversation turns."""
+
+import asyncio
+import struct
+
+import pytest
+
+from voice_assistant.agent.mock import MockAgentAdapter
+from voice_assistant.application.events import TurnCompleted, TurnOpened, TurnInterrupted, SpeechChunk
+from voice_assistant.application.voice_session import VoiceSession, VoiceSessionError
+from voice_assistant.domain.session import SessionState
+from voice_assistant.stt.mock import MockSTTAdapter
+from voice_assistant.tts.mock import MockTTSAdapter
+from voice_assistant.turn_manager import TurnManager
+from voice_assistant.vad.adapter import VADAdapter
+from voice_assistant.vad.models import SpeechStarted, SpeechEnded
+from voice_assistant.voice_pipeline import VoicePipeline
+
+
+@pytest.fixture
+def anyio_backend():
+    return "asyncio"
+
+
+class ScriptedVAD(VADAdapter):
+    def __init__(self, script=(), tail=()):
+        self.script = list(script)
+        self.tail = tail
+        self.received = []
+        self.closed = False
+
+    async def process(self, audio):
+        self.received.append(audio)
+        return self.script.pop(0) if self.script else ()
+
+    async def flush(self):
+        return self.tail
+
+    async def reset(self):
+        pass
+
+    async def aclose(self):
+        self.closed = True
+
+
+class RecordingSTT(MockSTTAdapter):
+    def __init__(self):
+        super().__init__()
+        self.received = []
+        self.aborted = False
+
+    async def send_audio(self, audio):
+        await super().send_audio(audio)
+        self.received.append(audio)
+
+    async def stop_session(self):
+        await self.emit_final("Olá")
+        await super().stop_session()
+
+    async def abort_session(self):
+        self.aborted = True
+        await super().abort_session()
+
+
+def build(vad, factory=RecordingSTT, event_sink=None, **options):
+    providers = []
+    events = []
+    completed = asyncio.Event()
+
+    def create_stt():
+        provider = factory()
+        providers.append(provider)
+        return provider
+
+    async def emit(event):
+        events.append(event)
+        if event_sink is not None:
+            await event_sink(event)
+        if isinstance(event, TurnCompleted):
+            completed.set()
+
+    session = VoiceSession(
+        pipeline=VoicePipeline(
+            session=SessionState(), turn_manager=TurnManager(),
+            agent=MockAgentAdapter(), tts=MockTTSAdapter(),
+        ),
+        create_stt=create_stt, emit=emit, vad=vad, **options,
+    )
+    return session, providers, events, completed
+
+
+def pcm(value, samples=256):
+    return struct.pack(f'<{samples}h', *([value] * samples))
+
+
+@pytest.mark.anyio
+async def test_delayed_start_replays_pre_roll_without_duplicating_audio():
+    vad = ScriptedVAD([(), (), (SpeechStarted(400),), ()])
+    session, providers, events, _ = build(vad, pre_roll_ms=10)
+    chunks = [pcm(i) for i in [1, 2, 3, 4]]
+    try:
+        await session.receive_audio(chunks[0])
+        await session.receive_audio(chunks[1])
+        assert providers == []
+        await session.receive_audio(chunks[2])
+        await session.receive_audio(chunks[3])
+        assert len(providers) == 1
+        assert b''.join(providers[0].received) == b''.join(chunks)[240 * 2:]
+        assert isinstance(events[0], TurnOpened)
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_speech_end_finalizes_in_background_while_silence_continues():
+    vad = ScriptedVAD([(SpeechStarted(0),), (SpeechEnded(400, 'silence'),), ()])
+    session, providers, events, completed = build(vad, pre_roll_ms=0)
+    try:
+        await session.receive_audio(pcm(1))
+        await session.receive_audio(pcm(2))
+        await session.receive_audio(pcm(3))
+        await asyncio.wait_for(completed.wait(), timeout=1.0)
+        assert b''.join(providers[0].received) == pcm(1) + pcm(2)[:144 * 2]
+        assert isinstance(events[-1], TurnCompleted)
+        assert session.active_turn_id is None
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_multiple_speech_intervals_in_one_chunk_keep_audio_separate():
+    vad = ScriptedVAD([(
+        SpeechStarted(0), SpeechEnded(160, 'silence'),
+        SpeechStarted(240), SpeechEnded(400, 'silence'),
+    )])
+    session, providers, events, completed = build(vad, pre_roll_ms=0)
+    audio = struct.pack('<512h', *range(512))
+    try:
+        await session.receive_audio(audio)
+        await asyncio.wait_for(completed.wait(), timeout=1.0)
+        assert len(providers) == 2
+        assert b''.join(providers[0].received) == audio[:320]
+        assert b''.join(providers[1].received) == audio[480:800]
+        assert any(isinstance(event, TurnInterrupted) and event.turn_id == 1 for event in events)
+        assert [event.turn_id for event in events if isinstance(event, TurnOpened)] == [1, 2]
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_new_speech_cancels_stt_finalization_and_opens_fresh_turn():
+    started = asyncio.Event()
+    cancelled = asyncio.Event()
+
+    class WaitingSTT(RecordingSTT):
+        async def stop_session(self):
+            started.set()
+            try:
+                await asyncio.Event().wait()
+            except asyncio.CancelledError:
+                cancelled.set()
+                raise
+
+    vad = ScriptedVAD([
+        (SpeechStarted(0),), (SpeechEnded(400, 'silence'),), (SpeechStarted(600),),
+    ])
+    session, providers, events, _ = build(vad, factory=WaitingSTT, pre_roll_ms=0)
+    try:
+        await session.receive_audio(pcm(1))
+        await session.receive_audio(pcm(2))
+        await asyncio.wait_for(started.wait(), timeout=1.0)
+        await session.receive_audio(pcm(3))
+        assert cancelled.is_set()
+        assert providers[0].aborted
+        assert len(providers) == 2
+        assert session.active_turn_id == 2
+        assert b''.join(providers[1].received) == pcm(3)[88 * 2:]
+        assert any(isinstance(event, TurnInterrupted) for event in events)
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+@pytest.mark.parametrize('audio', [b'', b'\x00', bytes(65538)])
+async def test_invalid_input_does_not_reach_vad_or_advance_history(audio):
+    vad = ScriptedVAD([(SpeechStarted(0),)])
+    session, providers, _, _ = build(vad, pre_roll_ms=0)
+    try:
+        with pytest.raises(ValueError):
+            await session.receive_audio(audio)
+        assert vad.received == []
+        await session.receive_audio(pcm(1))
+        assert b''.join(providers[0].received) == pcm(1)
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_manual_commands_cannot_compete_with_vad_turn_control():
+    session, _, _, _ = build(ScriptedVAD())
+    try:
+        for operation in [session.start_turn, lambda: session.send_audio(1, 0, pcm(1))]:
+            with pytest.raises(VoiceSessionError) as error:
+                await operation()
+            assert error.value.code == 'TURN_CONTROLLED_BY_VAD'
+        with pytest.raises(VoiceSessionError):
+            session.stop_turn(1)
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_input_error_prevents_unsafe_retry_and_cleanup_closes_vad():
+    class FailingVAD(ScriptedVAD):
+        async def process(self, audio):
+            raise RuntimeError('Inference failed')
+
+    vad = FailingVAD()
+    session, _, _, _ = build(vad)
+    try:
+        with pytest.raises(VoiceSessionError, match='Inference failed'):
+            await session.receive_audio(pcm(1))
+        with pytest.raises(VoiceSessionError, match='new session'):
+            await session.receive_audio(pcm(1))
+    finally:
+        await session.aclose()
+    assert vad.closed
+
+
+@pytest.mark.anyio
+async def test_speech_older_than_history_is_reported_instead_of_silently_truncated():
+    vad = ScriptedVAD([(), (SpeechStarted(0),)])
+    session, providers, _, _ = build(vad, pre_roll_ms=0, history_ms=1)
+    try:
+        await session.receive_audio(pcm(1))
+        with pytest.raises(VoiceSessionError, match='increase history_ms'):
+            await session.receive_audio(pcm(2))
+        assert providers == []
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_normal_end_of_input_flushes_active_speech():
+    vad = ScriptedVAD([(SpeechStarted(0),)], tail=(SpeechEnded(256, 'end_of_input'),))
+    session, _, _, completed = build(vad, pre_roll_ms=0)
+    try:
+        await session.receive_audio(pcm(1))
+        await session.finish_audio_input()
+        await asyncio.wait_for(completed.wait(), timeout=1.0)
+        with pytest.raises(VoiceSessionError, match='closed'):
+            await session.receive_audio(pcm(2))
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_session_end_during_vad_processing_does_not_open_late_turn():
+    started = asyncio.Event()
+    release = asyncio.Event()
+
+    class WaitingVAD(ScriptedVAD):
+        async def process(self, audio):
+            started.set()
+            await release.wait()
+            return (SpeechStarted(0),)
+
+    vad = WaitingVAD()
+    session, providers, events, _ = build(vad)
+    receiving = asyncio.create_task(session.receive_audio(pcm(1)))
+    ending = None
+    try:
+        await asyncio.wait_for(started.wait(), timeout=1.0)
+        ending = asyncio.create_task(session.end())
+        await asyncio.sleep(0)
+        release.set()
+        with pytest.raises(VoiceSessionError, match='shutdown'):
+            await receiving
+        await asyncio.wait_for(ending, timeout=1.0)
+        assert session.closed
+        assert providers == []
+        assert events == []
+    finally:
+        release.set()
+        await asyncio.gather(receiving, return_exceptions=True)
+        if ending:
+            await asyncio.gather(ending, return_exceptions=True)
+        await session.aclose()
+
+
+@pytest.mark.parametrize('options', [
+    {'pre_roll_ms': -1}, {'pre_roll_ms': True}, {'history_ms': 0},
+    {'history_ms': 100, 'pre_roll_ms': 200}, {'max_input_chunk_bytes': 0},
+])
+def test_invalid_audio_history_policy_is_rejected(options):
+    with pytest.raises(ValueError):
+        build(ScriptedVAD(), **options)
+
+
+@pytest.mark.anyio
+async def test_empty_stt_result_releases_turn_without_stopping_continuous_input():
+    interrupted = asyncio.Event()
+
+    async def sink(event):
+        if isinstance(event, TurnInterrupted):
+            interrupted.set()
+
+    vad = ScriptedVAD([
+        (SpeechStarted(0),), (SpeechEnded(400, 'silence'),), (SpeechStarted(600),),
+    ])
+    session, providers, _, _ = build(vad, factory=MockSTTAdapter, event_sink=sink)
+    try:
+        await session.receive_audio(pcm(1))
+        await session.receive_audio(pcm(2))
+        await asyncio.wait_for(interrupted.wait(), timeout=1.0)
+        assert session.active_turn_id is None
+        await session.receive_audio(pcm(3))
+        assert session.active_turn_id == 2
+        assert len(providers) == 2
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_new_speech_cancels_response_delivery_before_opening_next_turn():
+    delivering = asyncio.Event()
+    cancelled = asyncio.Event()
+
+    async def sink(event):
+        if isinstance(event, SpeechChunk) and event.turn_id == 1:
+            delivering.set()
+            try:
+                await asyncio.Event().wait()
+            except asyncio.CancelledError:
+                cancelled.set()
+                raise
+
+    vad = ScriptedVAD([
+        (SpeechStarted(0),), (SpeechEnded(400, 'silence'),), (SpeechStarted(600),),
+    ])
+    session, providers, events, _ = build(vad, event_sink=sink, pre_roll_ms=0)
+    try:
+        await session.receive_audio(pcm(1))
+        await session.receive_audio(pcm(2))
+        await asyncio.wait_for(delivering.wait(), timeout=1.0)
+        await session.receive_audio(pcm(3))
+        assert cancelled.is_set()
+        assert providers[0].aborted
+        assert session.active_turn_id == 2
+        opened = next(i for i, event in enumerate(events) if isinstance(event, TurnOpened) and event.turn_id == 2)
+        assert not any(isinstance(event, SpeechChunk) and event.turn_id == 1 for event in events[opened:])
+    finally:
+        await session.aclose()
+
+
+@pytest.mark.anyio
+async def test_pre_roll_never_replays_previous_utterance_into_next_stt():
+    vad = ScriptedVAD([(
+        SpeechStarted(0), SpeechEnded(160, 'silence'),
+        SpeechStarted(240), SpeechEnded(400, 'silence'),
+    )])
+    session, providers, _, completed = build(vad, pre_roll_ms=200)
+    audio = struct.pack('<512h', *range(512))
+    try:
+        await session.receive_audio(audio)
+        await asyncio.wait_for(completed.wait(), timeout=1.0)
+        assert b''.join(providers[0].received) == audio[:320]
+        assert b''.join(providers[1].received) == audio[320:800]
+    finally:
+        await session.aclose()