ONNX version
#1
by
PierreMesure
- opened
Hi,
I made an ONNX version of the model here in order 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?
I have two questions:
- would you like to add the ONNX weights to your own repositories so I can point to them? That way, people know what they're using and you centralise the download statistic on your repos.
- I tried the tiny models but the quality was much worse for me than the original OpenAI one with Finnish audio. Is there any reason for it? Your models are from 2023, do you know of any other experiment that got better results in Finnish transcription?
Was the Swedish conversion made from hf format or something different?
I maybe could look into ONNX conversion.
Most of our finetunes have improved performance from OpenAI models on Fleurs/Common Voice tests
I used this space to convert the tiny model. It uses the convert script provided with transformers.js.
You can use it locally by cloning the repo and running:
pip install -r scripts/requirements.txt
python -m scripts.convert --quantize --model_id Finnish-NLP/whisper-tiny-finnish