WhisperX missing pieces of transcript compared to Whisper API

[tomhayw]

I've been using WhisperX but I keep coming across issues whereby parts of the transcript are just missing entirely (i.e. half of sentences). I have ran the same audio file through OpenAI's Whisper API and it works perfectly fine.

Has anyone else had this issue and if so, how did you remediate it?

Thanks.

[sulutian]

You can try lowering --vad_onset 0.1 --vad_offset 0.1

[klausackermann]

I also obtained this and it is much better switching back to model large_v2 instead of model large_v3. The large_v3 also shows halo in the middle of transcribed, which large_v2 does not.