Sometimes segments with "no_speech_prob" larger than "speech threshold" and "avg_logprob" lower than "logprob_threshold" still appear about whisper-timestamped

[Jeronymous]

This is certainly not by design.

I double checked that no_speech_threshold and logprob_threshold are correctly passed to whisper.transcribe().

And with the options that you give (beam search...), the transcription is first done by whisper and segments are identical in the end (except from start and end timestamp). I mean there are little chance of bugs there (contrarily to the more tricky efficient case).

So I don't understand what can happen.
Can you please double check, and give what python call you make and an extract of the output that you find suspicious?
(Also you can test what gives whisper alone on this)

[ItakeLs]

Yes, you are right, the issue was on my end, after thorough checking, it is working as intended.