fix(voice): playback-aware muting when \--speak\ and \--voice\ run to…

CHANGELOG.md

@@ -9,6 +9,10 @@ All notable changes to this project are documented here. The format is based on
 
 ### Fixed
 
+- **Voice plugin (0.5.2):** added playback-aware muting when `--speak` and
+  `--voice` run together, zero-filling captured microphone blocks during active
+  TTS playback ([#332](https://github.com/robocurve/inspect-robots/issues/332)).
+
 - **Voice plugin (0.5.1):** operator-ended trials now cut `--speak` narration
   instead of draining it at eval end
   ([plan 0061](plans/0061-speak-operator-end-cut.md),

plugins/inspect-robots-voice/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "inspect-robots-voice"
-version = "0.5.1"
+version = "0.5.2"
 description = "Local spoken operator feedback for attended Inspect Robots evaluations."
 dynamic = ["readme"]
 requires-python = ">=3.10"

plugins/inspect-robots-voice/src/inspect_robots_voice/__init__.py

@@ -14,7 +14,7 @@
 
 __all__ = ["SpeakerSink", "VoiceInput", "speaker_sink", "voice_input"]
 
-__version__ = "0.5.1"
+__version__ = "0.5.2"
 
 ScalarValue = str | int | float | bool | None
 

plugins/inspect-robots-voice/src/inspect_robots_voice/_capture.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import queue as queue_module
+import threading
 import warnings
 from collections.abc import Mapping, Sequence
 from contextlib import suppress
@@ -15,6 +16,10 @@
 Device = str | int | None
 
 
+_speakers_lock = threading.Lock()
+_active_speakers: set[object] = set()
+
+
 class MicrophoneCapture:
     """Capture 16-bit-equivalent mono float blocks without blocking PortAudio's callback."""
 
@@ -130,6 +135,10 @@ def _callback(
     ) -> None:
         del frames, time_info, status
         block = np.asarray(indata, dtype=np.float32).reshape(-1).copy()
+        with _speakers_lock:
+            active = bool(_active_speakers)
+        if active:
+            block.fill(0)
         try:
             self._queue.put_nowait(block)
             return

plugins/inspect-robots-voice/src/inspect_robots_voice/_speaker.py

@@ -309,12 +309,20 @@ def _worker(self, engine: TtsEngine, playback: _Playback) -> None:
                     continue
                 gained = np.asarray(samples * np.float32(self.volume), dtype=np.float32)
                 chunk_size = max(1, int(sample_rate * _CHUNK_SECONDS))
-                for start in range(0, len(gained), chunk_size):
-                    if self._stop.is_set():
-                        return
-                    if self._speech_gen != gen:
-                        break
-                    playback.write(gained[start : start + chunk_size], sample_rate)
+                from inspect_robots_voice._capture import _active_speakers, _speakers_lock
+
+                with _speakers_lock:
+                    _active_speakers.add(self)
+                try:
+                    for start in range(0, len(gained), chunk_size):
+                        if self._stop.is_set():
+                            return
+                        if self._speech_gen != gen:
+                            break
+                        playback.write(gained[start : start + chunk_size], sample_rate)
+                finally:
+                    with _speakers_lock:
+                        _active_speakers.discard(self)
             except Exception as exc:
                 with self._condition:
                     self._disabled = True

plugins/inspect-robots-voice/tests/test_capture.py

@@ -172,3 +172,29 @@ def failing_import(name: str, *args: object, **kwargs: object) -> object:
 
     with pytest.raises(OSError, match="libportaudio2"):
         MicrophoneCapture(None, 16_000, audio_queue)
+
+
+def test_playback_aware_muting_callback() -> None:
+    from inspect_robots_voice._capture import _active_speakers, _speakers_lock
+
+    sounddevice = _SoundDevice(_DEVICES)
+    audio_queue: queue.Queue[np.ndarray] = queue.Queue(maxsize=2)
+    capture = MicrophoneCapture(1, 16_000, audio_queue, _sounddevice=sounddevice)
+
+    # When no speakers are active, blocks should be preserved.
+    with _speakers_lock:
+        _active_speakers.clear()
+
+    capture._callback(np.array([[5.0], [6.0]]), 2, object(), object())
+    assert np.array_equal(audio_queue.get_nowait(), np.array([5.0, 6.0], dtype=np.float32))
+
+    # When a speaker is active, blocks should be zero-filled.
+    dummy_speaker = object()
+    with _speakers_lock:
+        _active_speakers.add(dummy_speaker)
+    try:
+        capture._callback(np.array([[5.0], [6.0]]), 2, object(), object())
+        assert np.array_equal(audio_queue.get_nowait(), np.array([0.0, 0.0], dtype=np.float32))
+    finally:
+        with _speakers_lock:
+            _active_speakers.discard(dummy_speaker)

plugins/inspect-robots-voice/tests/test_factory.py

@@ -11,7 +11,7 @@
 
 
 def test_package_exports_and_version() -> None:
-    assert inspect_robots_voice.__version__ == "0.5.1"
+    assert inspect_robots_voice.__version__ == "0.5.2"
     assert inspect_robots_voice.__all__ == [
         "SpeakerSink",
         "VoiceInput",

plugins/inspect-robots-voice/tests/test_speaker.py

@@ -748,3 +748,35 @@ def enqueue_degraded() -> None:
 
     playback.release[0].set()
     sink.close()
+
+
+def test_speaker_active_playback_registration() -> None:
+    from inspect_robots_voice._capture import _active_speakers, _speakers_lock
+
+    engine = _FakeEngine()
+    playback = _GatedPlayback(gated_writes=1)
+    sink = _sink(engine, playback)
+    sink.start()
+
+    with _speakers_lock:
+        _active_speakers.clear()
+
+    sink.log_policy_messages(
+        0, [_assistant(_tool_call("move", {"note": "test-playback-registration"}))]
+    )
+
+    # Wait for the worker to synthesize and start writing the chunk
+    assert playback.entered[0].wait(timeout=2.0)
+
+    # The speaker should be registered in _active_speakers during playback
+    with _speakers_lock:
+        assert sink in _active_speakers
+
+    # Release the playback write chunk
+    playback.release[0].set()
+    _wait_until(lambda: len(playback.writes) == 3)
+    sink.close()
+
+    # The speaker should be discarded from _active_speakers when done
+    with _speakers_lock:
+        assert sink not in _active_speakers

plugins/inspect-robots-voice/tests/test_tts.py

@@ -42,7 +42,9 @@ def test_missing_explicit_path_names_the_path(
     missing = tmp_path / "missing"
     monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "cache"))
 
-    with pytest.raises(FileNotFoundError, match=str(missing)):
+    import re
+
+    with pytest.raises(FileNotFoundError, match=re.escape(str(missing))):
         resolve_model_files(
             str(missing) if kind == "model" else str(present),
             str(missing) if kind == "voices" else str(present),