Better Diarization pipeline

[omarsiddiqi224]

For some audio files, the diarization works, while for others they do not. If I run the audio file that didn't work with only transcription, no diarization, then it works perfectly. I have attached 2 screenshots showcasing the difference.

Audio file that works:

Audio file that does not work:

I believe the issue is in diarize.py. I found someone with the same error in another repo as well (https://github.com/huggingface/speechbox/issues/28), but their solution did not work for me. A person did try to solve it, but the repo is different. Here is his solution (https://github.com/huggingface/speechbox/pull/35).

[Vaibhavs10]

Hey @omarsiddiqi224 - I just released a patch. Which should fix this. Can you please run the following:

1. pipx uninstall insanely-fast-whisper
2. pipx install insanely-fast-whisper==0.0.13 --force

[oliverwehrens]

Thanks a lot. I had the same problem and it seemed to have fixed the problem.

[omarsiddiqi224]

Thank you, yes it worked

[Pikauba]

Be careful about this so called Fix. As it is the same exact code used in speechbox (I wonder why the speechbox library is not directly integrated in this repo instead?) the speakers assigned won't be the good one for the reasons I mentioned in the above cited issue.

This workaround just ignore an underlying problem and for an audio long enough the results will be totally wrong. But that is my 2 cents.

[Vaibhavs10]

Hey hey @Pikauba - The reason I added it over here is so that we can maintain it better, since, Speechbox isn't actively maintained at the moment.

Also, note that we will be rolling out a new production-ready whisper package in January.

I'd be open to add it here, if you would like to raise a similar PR here as well!

Thank you for your brilliant contribution! 🤗

[FredHaa]

Any word on when we can expect this new whisper package?