ONNX version

#2
by PierreMesure - opened

Hi, awesome work with this Icelandic model!

I'd like to make it available at whisper-web.mesu.re.

It's an app that uses transformers.js for local inference in the browser. It's proven very useful in Sweden, being the easiest way to use the recently KB-whisper finetunes and I'm trying to add other quality finetunes like nb-whisper and maybe yours?

In order to do so, I would need an ONNX version of the model. It can be stored on your repo or anywhere else. I've converted the Norwegian models here, for example.

I have three questions:

  • Would you like to add the ONNX weights to your own repository so I can point to them? That way, people know what they're using and you centralise the download statistic on your repo.
  • Have you produced smaller versions like a tiny, base, small, medium? In the browser or on mobile, smaller variants are very useful
  • Do you have any performance comparisons between OpenAI model and this fine-tune?
Language and Voice Laboratory (Reykjavík University) org

Dear Pierre, how are you?
Thank you very much for your message. Regarding to your questions:

  • Yes, there is no problem in adding the ONNX weights in our repo. Thaks for giving us the option.
  • Unfortunately, we don't have smaller versions of this model. You can find some smaller versions in my personal repo https://huggingface.co/carlosdanielhernandezmena but they are trained with other data.
  • We don't have a comparison performance. Actually, I remember that the OpenAI's Whisper was not good in Icelandic, that is why we had to finetune a model to validate the data that we used in the model that you want to use. We report that in this paper: https://aclanthology.org/2024.lrec-main.1246.pdf

Hope this helps.
Please let me know how we can coordinate to upload the weights.

Best Regards,
Carlos

Sign up or log in to comment