Comparision with original whisper timestamps?

[naymaraq]

Hi,

I would like to test timestamps algorithm by calculating Detection Error Rate. I saw in the original Whisper-repo, that there is an command line option --word-timestamps. I am wondering is anyone compare these two implementations (whisper-timestamped and original whisper). And how they differ?

[Jeronymous]

We did some comparison which show that timestamps of whisper-timestamped tend to be a bit more accurate than the ones in original whisper.

Those results were not fully published, but there are some details in : "Transcribing And Aligning Conversational Speech: A Hybrid Pipeline Applied To French Conversations" (ASRU 2023)
Precision of openai-whisper in terms of timestamps is similar to the one of faster_whisper (that re-implements the same algorithm).

A direct comparison of openai-whisper and whisper-timestamped is difficult, because whisper-timestamps offers additional options such as the VAD (which is critical to obtain good results).