Whisper-v3

[suryasanchez]

Update to support Whisper v3? openai/whisper#1762

[thomasmol]

Yes I am working on it. Currently there are issues converting whisper v3 to cttranslate SYSTRAN/faster-whisper#544 (comment)

[thomasmol]

Currently waiting on this PR to get merged: huggingface/transformers#26699 so we can use batched inference and still get word level timestamps. I'll move away from faster-whisper (the maintainer is now working at Apple and not actively maintaining the repo anymore) and use the hugginface/transformer implementation described here: https://huggingface.co/openai/whisper-large-v3. Should result in even faster inference over faster-whisper. Hopefully somewhere this or next week!

[thomasmol]

Faster-whisper 0.10.0 was released, which includes support for whisper v3. Faster-whisper also has a new maintainer and moved here: https://github.com/SYSTRAN/faster-whisper.
Just updated the pipeline to new version of faster-whisper and to new version of pyannote (3.1).
Will still look into the batched inference when that's possible, but only when it supports VAD and word level timestamps.