wx44wx/pipecat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Don't inject silence in the proxy.

Browse Source
This commit is contained in:
filipi87
2026-05-20 18:26:11 -03:00
parent 0dc1337a3c
commit 1338da6831
1 changed files with 6 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
scripts/daily/test_tavus_transport.py
View File

@@ -61,7 +61,7 @@ class DailyProxyApp(EventHandler):
) )
# Playback source declared at TRUE_SAMPLE_RATE — consumes audio at real-time speed. # Playback source declared at TRUE_SAMPLE_RATE — consumes audio at real-time speed.
self._audio_source = CustomAudioSource(TRUE_SAMPLE_RATE, 1) self._audio_source = CustomAudioSource(TRUE_SAMPLE_RATE, 1, False)
self._audio_track = CustomAudioTrack(self._audio_source) self._audio_track = CustomAudioTrack(self._audio_source)
def on_joined(self, data, error): def on_joined(self, data, error):
@@ -162,16 +162,11 @@ class DailyProxyApp(EventHandler):
async def _buffer_audio(self, audio_data: AudioData): async def _buffer_audio(self, audio_data: AudioData):
"""Append received bytes to the buffer, skipping WebRTC-injected silence. """Append received bytes to the buffer, skipping WebRTC-injected silence.
Speech frames arrive at DECLARED_SAMPLE_RATE speed (~2x real-time) and Speech frames arrive at DECLARED_SAMPLE_RATE speed (~2x real-time) so the
are added in full — the buffer grows ahead of playback, which is the buffer grows ahead of the drain. WebRTC-injected silence (all-zero PCM) is
fast-audio effect we want to observe. The buffer then drains at handled differently based on buffer level: below MIN_AUDIO_BUFFER we keep it
TRUE_SAMPLE_RATE speed once speech stops. so the pre-buffer can fill; above that threshold we discard it so the buffer
drains back down between utterances.
On the first silence after speech (speech→silence transition) we insert
one real-time-equivalent chunk of silence to represent the natural pause.
All subsequent silence frames are dropped so the buffer drains back down.
This also limits the impact of any TTS-emitted silence to a single chunk
per transition rather than letting it accumulate.
""" """
new_bytes = audio_data.audio_frames new_bytes = audio_data.audio_frames
if self._is_silence(new_bytes): if self._is_silence(new_bytes):