How to access the internal vad timestamps?

[at-src]

I want to access whatever timestamp silero vad produces internally based on which it gives timestamp if vad fileter is true. I tried get_speech_timestamps() method but it seems this requires preprocessing the audio again and i think it recalculates based on what i provide rather than what faster whisper uses. Is there any way to access it? I even tried recreating the entire vad enabled pipeline but no luck.

[arcticloud]

get_speech_timestamps() isn't recomputing something different — it's the exact function faster-whisper calls internally. The reason your output didn't match is almost certainly a different sample rate or different VadOptions than the ones transcribe() uses. Feed it the same 16 kHz decoded audio and the same options and you reproduce FW's internal segments exactly:

from faster_whisper.audio import decode_audio
from faster_whisper.vad import get_speech_timestamps, VadOptions

audio = decode_audio("audio.wav", sampling_rate=16000)   # exactly what FW feeds the VAD
vad_opts = VadOptions()                                   # must match what you pass to transcribe(vad_parameters=...)
speech = get_speech_timestamps(audio, vad_opts)
# -> [{'start': <samples>, 'end': <samples>}, ...]

for s in speech:
    print(s["start"] / 16000, s["end"] / 16000)          # samples -> seconds

Key points: always decode at 16000 Hz (the VAD is fixed to that rate), and if you customized vad_parameters in transcribe(), pass the identical values into VadOptions(...) here. Do that and the timestamps are bit-for-bit the ones FW used internally to filter the audio.