New / co maintainer for faster whisper

AvivSham · November 9, 2023, 10:40am

Hi All,
How are you?
Thank you for your amazing work and contribution to the ML/DL community!
The repo faster-whisper is heavily rely on Ctranslate. Given the fact that Guillaume Klein is the maintainer of faster-whisper and he started a new position that consumes most of his time, I think we need a new / co maintainer for faster-whisper. New Whisper versions will be released and code adaptations will be needed to support these versions.
Can you please assign someone for this important mission?

Thanks in advance!

vince62s · November 9, 2023, 12:45pm

github.com/guillaumekln/faster-whisper

Support for Distil-Whisper

opened 05:06PM - 01 Nov 23 UTC

sanchit-gandhi

Hey @guillaumekln! Thanks for this fantastic resource. We're looking at supporti…ng the [Distil-Whisper](https://github.com/huggingface/distil-whisper) checkpoints in faster-whisper. The checkpoints are fairly easy to convert: we just pin the number of decoder layers to 2 always, and load the 32 encoder layers. For inference, we found a chunk length of 15-20s to be optimal for WER performance of the distilled model, see Table 23 of the [paper](https://arxiv.org/abs/2311.00430): <img width="858" alt="Screenshot 2023-11-01 at 17 05 58" src="https://github.com/guillaumekln/faster-whisper/assets/93869735/ea4ea9b6-465b-4344-8663-1dc92abcf6b2"> Would you be open to a PR allowing the user to specify the chunk length and also the maximum generation length? This would enable full support of Distil-Whisper in faster-whisper!