WhisperX implementation

[JonasCz]

This adds WhisperX, useful to me because it's faster and has word-level timestamps.

(Word level timestamps are used here in order to "regroup" subtitle lines more logically, e.g. each sentence gets its own subtitle line if possible, and otherwise lines are split in a more sensible way instead of the pretty "random" output of base Whisper).

This probably:

Closes (sort of) #53
Progress #58 (WhisperX supports diarization, although this doesn't do that)

In order to use it, you have to:

• Add WhisperX to transcription-api/requirements.txt instead of faster-whisper
• Replace call to FasterWhisperBackend in transcribe.py with WhisperxBackend

The code is quite messy, it's "minimal effort" adapted from one of my other projects.

[pluja]

Hey, thanks for this. I think I will work based on your implementation on a better integration. Also, note that Faster-Whisper also has word-level timestamps, but Whishper is not using them (although if you get the JSON subtitles you'll see the word-level distinction).

The idea is that the API should be able to support multiple backends so that the user can choose between WhisperX and Faster-Whisper.

I think I will be experimenting with this, but if WhisperX is better than Faster-Whisper as it seems, and can also run with CPU, maybe it would be better to switch to it by default instead of faster-whisper.

I'll be looking at this and come back here when I have experimented with it.

I'd also like to hear what you (and any other user) thinks!

[wobsoriano]

Would love this!

[SandiyosDev]

@pluja will this be merged?

[pluja]

Hi @SandiyosDev if you are interested in a WhisperX implementation, you can try the WhisperX-API standalone of the new (to be finished) Whishper-v4 rewrite (will be renamed to Anysub):

https://github.com/pluja/whishper/blob/v4/whisperx-api%2FREADME.md