Word-level timestamps are very inaccurate

[a-rogalska]

I'm using large-v2 model to transcribe multilingual audio (many of them are in German). There are many cases, usually at the beginning of the segment, when word-level timestamps are incorrect, with the start time later than the end time. I know that whisper-timestamped has pretty accurate results, but I would like to use faster-whisper instead of the original whisper implementation.

Is there a way to improve timestamp accuracy here?

[Purfview]

Try #226 PR, it helps with word timestamps.

[a-rogalska]

It helped, thanks.