Degraded quality with timestamps disabled

[jpospical]

After commit f7908f9, some audio is often ignored when running with -nt. This seems to happen regardless of audio file, model or CPU (I've tested with Intel and AMD at least). I can reliably reproduce with this clip and large-v1. Specifically, all but the last word "damage" of the section from 00:00:27-00:00:30 is consistently ignored with timestamps=0, but it's transcribed as expected otherwise.

I was going to comment on #1755 rather than creating a new issue, but that PR was merged a long time ago so I wasn't sure if it was appropriate. Apologies if that's what I should've done.

[au-voltzzz]

I commented on a different thread ( #2191) which I thought described my issue at first but just inspired me to try the earlier version of whisper.cpp. However, I have just discovered using -nt was the culprit of my issues after trying different things thorughout the day. It happened using v2, v3, and medium.en models. Large portions of audio were not transcribed. For now I am using whisper.cpp-v1.5.4 which fixes the problem.

This is happening for me using Apple M1 (sorry I am not sure what info to share to help describe my issue better)

Edit: I removed the comment on the other thread

[Macbit]

I’m experiencing a similar issue on v1.7.1 running on M3. When I remove timestamps from the transcript, some words end up repeating 20+ times. However, using the same source and settings with timestamps works perfectly.