Skip to content

fix: improve vad sensitivity and buffer size for better transcription accuracy #18

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
MusicMaster4 merged 1 commit into from
Aug 1, 2026
Merged

fix: improve vad sensitivity and buffer size for better transcription accuracy #18

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 14 additions & 7 deletions python/dictation_service.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,7 +160,9 @@ def __init__(self) -> None:
self.frame_ms = 30
self.frame_samples = int(self.sample_rate * self.frame_ms / 1000)
self._init_spectrum_bands()
self.vad = webrtcvad.Vad(2)
# Prefer retaining marginal speech over aggressively removing it. Whisper can
# ignore a little background noise, but it cannot recover clipped consonants.
self.vad = webrtcvad.Vad(1)
self.model_name = os.getenv("WHISPER_MODEL", "small")
self.cloud_mode = os.getenv("FLOW_TRANSCRIPTION_ENGINE", "local").lower() == "cloud"
self.model_dir = os.getenv("WHISPER_MODEL_DIR")
Expand All @@ -176,7 +178,9 @@ def __init__(self) -> None:
self.stream: Optional[sd.InputStream] = None
self.listening = False
self.triggered = False
self.ring_buffer = deque(maxlen=8)
# Keep enough audio before and after VAD activation to preserve quiet word
# beginnings and endings, which are especially important for proper nouns.
self.ring_buffer = deque(maxlen=12)
self.voiced_frames = []
self.silence_frames = 0
self.pending_segments: list[np.ndarray] = []
Expand Down Expand Up @@ -342,11 +346,14 @@ def _transcribe_segment(self, segment: np.ndarray, language: str):
return self.model.transcribe(
segment,
language=language,
beam_size=5,
best_of=5,
beam_size=8,
best_of=8,
patience=1.2,
vad_filter=False,
condition_on_previous_text=False,
temperature=0.0,
condition_on_previous_text=True,
# Start deterministically, then retry only difficult passages with a
# small amount of sampling instead of accepting a weak first decode.
temperature=(0.0, 0.2, 0.4),
compression_ratio_threshold=2.4,
no_speech_threshold=0.45,
)
Expand Down Expand Up @@ -634,7 +641,7 @@ def _process_frame(self, frame: bytes) -> None:
self.voiced_frames.append(frame)
self.silence_frames = 0 if is_speech else self.silence_frames + 1

if self.silence_frames >= 12 or len(self.voiced_frames) >= self.max_segment_frames:
if self.silence_frames >= 18 or len(self.voiced_frames) >= self.max_segment_frames:
frames = self.voiced_frames
self._reset_segment_state()
if len(frames) >= self.min_segment_frames:
Expand Down
Loading