admin-valentin/WhisperLiveKit
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Improve tokens alignment and silence handling

Browse source
This commit is contained in:
Quentin Fuxa 2026-01-24 10:55:00 +01:00
parent 6e85c16614
commit a282cbe75f
1 changed files with 26 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

36
whisperlivekit/tokens_alignment.py
View file

@ -1,9 +1,15 @@
from time import time from time import time
from typing import Any, List, Optional, Tuple, Union from typing import Any, List, Optional, Tuple, Union
from whisperlivekit.timed_objects import (ASRToken, Segment, PuncSegment, Silence, from whisperlivekit.timed_objects import (
SilentSegment, SpeakerSegment, ASRToken,
TimedText) PuncSegment,
Segment,
Silence,
SilentSegment,
SpeakerSegment,
TimedText,
)
_DEFAULT_RETENTION_SECONDS: float = 300.0 _DEFAULT_RETENTION_SECONDS: float = 300.0
@ -193,7 +199,7 @@ class TokensAlignment:
max_overlap = intersec max_overlap = intersec
max_overlap_speaker = diarization_segment.speaker + 1 max_overlap_speaker = diarization_segment.speaker + 1
punctuation_segment.speaker = max_overlap_speaker punctuation_segment.speaker = max_overlap_speaker
segments = [] segments = []
if punctuation_segments: if punctuation_segments:
segments = [punctuation_segments[0]] segments = [punctuation_segments[0]]
@ -209,12 +215,22 @@ class TokensAlignment:
def get_lines( def get_lines(
self, self,
diarization: bool = False, diarization: bool = False,
translation: bool = False, translation: bool = False,
current_silence: Optional[Silence] = None current_silence: Optional[Silence] = None,
audio_time: Optional[float] = None,
) -> Tuple[List[Segment], str, Union[str, TimedText]]: ) -> Tuple[List[Segment], str, Union[str, TimedText]]:
"""Return the formatted segments plus buffers, optionally with diarization/translation.""" """Return the formatted segments plus buffers, optionally with diarization/translation.
Args:
audio_time: Current audio stream position in seconds. Used as fallback
for ongoing silence end time instead of wall-clock (which breaks
when audio is fed faster or slower than real-time).
"""
# Fallback for ongoing silence: prefer audio stream time over wall-clock
_silence_now = audio_time if audio_time is not None else (time() - self.beg_loop)
if diarization: if diarization:
segments, diarization_buffer = self.get_lines_diarization() segments, diarization_buffer = self.get_lines_diarization()
else: else:
@ -225,7 +241,7 @@ class TokensAlignment:
self.validated_segments.append(Segment.from_tokens(self.current_line_tokens)) self.validated_segments.append(Segment.from_tokens(self.current_line_tokens))
self.current_line_tokens = [] self.current_line_tokens = []
end_silence = token.end if token.has_ended else time() - self.beg_loop end_silence = token.end if token.has_ended else _silence_now
if self.validated_segments and self.validated_segments[-1].is_silence(): if self.validated_segments and self.validated_segments[-1].is_silence():
self.validated_segments[-1].end = end_silence self.validated_segments[-1].end = end_silence
else: else:
@ -235,13 +251,13 @@ class TokensAlignment:
)) ))
else: else:
self.current_line_tokens.append(token) self.current_line_tokens.append(token)
segments = list(self.validated_segments) segments = list(self.validated_segments)
if self.current_line_tokens: if self.current_line_tokens:
segments.append(Segment.from_tokens(self.current_line_tokens)) segments.append(Segment.from_tokens(self.current_line_tokens))
if current_silence: if current_silence:
end_silence = current_silence.end if current_silence.has_ended else time() - self.beg_loop end_silence = current_silence.end if current_silence.has_ended else _silence_now
if segments and segments[-1].is_silence(): if segments and segments[-1].is_silence():
segments[-1] = SilentSegment(start=segments[-1].start, end=end_silence) segments[-1] = SilentSegment(start=segments[-1].start, end=end_silence)
else: else: