c9ea2140ad
AudioRecorder.stop() applied its final too-quiet discard with the hardcoded SILENCE_RMS_THRESHOLD (200) instead of self._silence_threshold, which the CLI wires from voice.silence_threshold before each recording. On a low-threshold setup (e.g. 80 for a quiet mic that peaks at RMS ~160) the VAD correctly detected speech, auto-stopped, and then stop() discarded the valid recording as too quiet — every turn after the first failed with "No speech detected". The streaming-TTS arm check in _chat_setup_turn_audio also probed _import_sounddevice() before check_tts_requirements(); on macOS that raises (TCC media-library prompt, which is why the speaker path deliberately avoids sounddevice on Darwin and routes output via tempfile -> afplay), and the bare except silently disabled streaming TTS for every macOS voice user. stop() now applies the same configured floor as the VAD, and arming is check_tts_requirements() alone; the arm failure is logged at debug instead of being swallowed. Co-authored-by: Professor Dombili <213498199+CryptoDombili@users.noreply.github.com>
96 lines
3.9 KiB
Python
96 lines
3.9 KiB
Python
"""Bug-fix regression tests for #84046: ``AudioRecorder.stop()`` must apply the CONFIGURED
|
|
silence threshold (``self._silence_threshold``, wired from ``voice.silence_threshold`` by the
|
|
CLI before each recording), not the hardcoded ``SILENCE_RMS_THRESHOLD`` default of 200 —
|
|
otherwise a low-threshold setup (e.g. 80 for a quiet mic that peaks at RMS ~160) auto-stops on
|
|
speech and then discards the valid recording as "too quiet".
|
|
|
|
numpy is absent from the test environment, so captured frames are faked as dicts that carry
|
|
just the attributes ``stop()`` touches (``__len__``, concatenation). This is a behavior
|
|
contract on the threshold, not on any waveform shape.
|
|
"""
|
|
|
|
import pytest
|
|
|
|
from tools.voice_mode import AudioRecorder, SILENCE_RMS_THRESHOLD
|
|
|
|
|
|
class _Frame:
|
|
"""Minimal frame stand-in: length + concatenation + bytes, everything stop()
|
|
and _write_wav touch. A constant-PCM stand-in would need numpy to build;
|
|
plain zero bytes exercise the write path the same way the old silence
|
|
fixture did."""
|
|
|
|
def __init__(self, n):
|
|
self._n = n
|
|
|
|
def __len__(self):
|
|
return self._n
|
|
|
|
def tobytes(self):
|
|
return b"\x00\x00" * self._n
|
|
|
|
|
|
class _FakeAudio:
|
|
def concatenate(self, frames, axis=0):
|
|
assert axis == 0
|
|
assert len(frames) == 1
|
|
return frames[0]
|
|
|
|
|
|
@pytest.fixture
|
|
def mock_audio(monkeypatch, tmp_path):
|
|
monkeypatch.setattr("tools.voice_mode._import_audio", lambda: (None, _FakeAudio()))
|
|
monkeypatch.setattr("tools.voice_mode._TEMP_DIR", str(tmp_path))
|
|
return _FakeAudio()
|
|
|
|
|
|
def _arm(recorder, *, frames, peak_rms):
|
|
"""Put the recorder into the state a live capture leaves behind, without
|
|
opening a real InputStream: recording flag on, frames + peak RMS loaded."""
|
|
recorder._recording = True
|
|
recorder._frames = list(frames)
|
|
recorder._peak_rms = peak_rms
|
|
|
|
|
|
class TestStopUsesConfiguredSilenceThreshold:
|
|
|
|
def test_stop_keeps_recording_above_configured_threshold(self, mock_audio):
|
|
"""peak RMS above the CONFIGURED threshold (but below the 200 default) is
|
|
valid speech and must be returned, not discarded as too quiet."""
|
|
recorder = AudioRecorder()
|
|
recorder._silence_threshold = 80 # the user's voice.silence_threshold
|
|
_arm(recorder, frames=[_Frame(16000)], peak_rms=160) # above 80, below 200
|
|
|
|
assert recorder.stop() is not None
|
|
|
|
def test_stop_discards_recording_below_configured_threshold(self, mock_audio):
|
|
"""peak RMS below the CONFIGURED threshold is still silence and must be
|
|
discarded — lowering the floor never lets noise through."""
|
|
recorder = AudioRecorder()
|
|
recorder._silence_threshold = 80
|
|
_arm(recorder, frames=[_Frame(16000)], peak_rms=40) # below 80
|
|
|
|
assert recorder.stop() is None
|
|
|
|
def test_stop_threshold_tracks_reconfigured_threshold_between_recordings(self, mock_audio):
|
|
"""The discard gate follows the live attribute: the same peak RMS that was
|
|
discarded under threshold 80 is kept after the CLI re-wires it to 60 before
|
|
the next recording (per-recording reconfiguration, cli_voice_mixin)."""
|
|
recorder = AudioRecorder()
|
|
recorder._silence_threshold = 80
|
|
_arm(recorder, frames=[_Frame(16000)], peak_rms=70)
|
|
assert recorder.stop() is None # below 80: discarded
|
|
|
|
recorder._silence_threshold = 60
|
|
_arm(recorder, frames=[_Frame(16000)], peak_rms=70)
|
|
assert recorder.stop() is not None # above 60: kept
|
|
|
|
def test_stop_still_discards_below_default_threshold_without_config(self, mock_audio):
|
|
"""With no config override the constructor default (200) still applies: RMS
|
|
160 stays a silent recording. Pins the default against the custom floor."""
|
|
recorder = AudioRecorder()
|
|
assert recorder._silence_threshold == SILENCE_RMS_THRESHOLD
|
|
_arm(recorder, frames=[_Frame(16000)], peak_rms=160)
|
|
|
|
assert recorder.stop() is None
|