pepperplus/pepperplus-cb
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

feat: prepend more silence before speech audio for better transcription beginnings

Browse Source 
ref: N25B-429
This commit is contained in:
Twirre Meulenbelt
2026-01-08 10:49:13 +01:00
parent 5a61225c6f
commit 45719c580b
3 changed files with 10 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/control_backend/agents/perception/vad_agent.py
View File

@@ -229,10 +229,11 @@ class VADAgent(BaseAgent):
assert self.model is not None
prob = self.model(torch.from_numpy(chunk), settings.vad_settings.sample_rate_hz).item()
non_speech_patience = settings.behaviour_settings.vad_non_speech_patience_chunks
begin_silence_length = settings.behaviour_settings.vad_begin_silence_chunks
prob_threshold = settings.behaviour_settings.vad_prob_threshold
if prob > prob_threshold:
if self.i_since_speech > non_speech_patience:
if self.i_since_speech > non_speech_patience + begin_silence_length:
self.logger.debug("Speech started.")
self.audio_buffer = np.append(self.audio_buffer, chunk)
self.i_since_speech = 0
@@ -246,11 +247,12 @@ class VADAgent(BaseAgent):
continue
# Speech probably ended. Make sure we have a usable amount of data.
if len(self.audio_buffer) >= 3 * len(chunk):
if len(self.audio_buffer) > begin_silence_length * len(chunk):
self.logger.debug("Speech ended.")
assert self.audio_out_socket is not None
await self.audio_out_socket.send(self.audio_buffer[: -2 * len(chunk)].tobytes())
# At this point, we know that the speech has ended.
# Prepend the last chunk that had no speech, for a more fluent boundary
self.audio_buffer = chunk
# At this point, we know that there is no speech.
# Prepend the last few chunks that had no speech, for a more fluent boundary.
self.audio_buffer = np.append(self.audio_buffer, chunk)
self.audio_buffer = self.audio_buffer[-begin_silence_length * len(chunk) :]