Very impressed.... but...

[fraternl]

I used this software to create a subtitle for "The Princes in the Tower" with the program Subtitleedit which uses Whisper as its STT-engine.
I was in awe how accurate it created the text for this movie. It's the only time I used it thus far, because it was alas still a lot of work to get a good subtitle, but this wasn't for the text it generated.

It actually needed only a few corrections.

No, the problem was only with the timing.

I totally do not comprehend why a very difficult task as converting spoken text to written text can be done, but then having the timing wrong on so many places.

I do not know how Whisper is integrated into Subtitleedit, but when I asked the programmer, he said it was the only weak point of Whisper.

Is this timing problem recognised?

BTW... I used the CPP-engine and I haven't tried any others, yet.

[ulatekh]

Stock Whisper doesn't generate accurate timestamps; more often than not, they're only accurate to the nearest whole second. Derived projects like WhisperX employ additional AI models (e.g. wav2vec v2) to improve timestamp accuracy.

[clort81]

I can get whisper.cpp to run.
I can't get WhisperX to run.

Python ecosystem is a dumpster fire. Or something like purgatory.

Blessings to ggerganov!