Smaller Whisper model?

[DigitLib]

Tested to convert fine tuned model and use it in Whisper. Work excellent! Thank you!

Whisper model is 2.9G while in HF is 6.17G.
Use this fine tuned model
https://huggingface.co/DrishtiSharma/whisper-large-v2-serbian/tree/main
Also 6.17G but get 5.7G.
Q: is there a way to reduce size to 2.9G for example? To be easier to download.

Use it in https://github.com/DigitLib/whisper-webui-vad

Just add new large-v2-sr and works excellent in Whisper app!!!

...and Happy New Year!!!

[bayartsogt-ya]

Hey @DigitLib , sorry for the slow response.

Sounds so cool that you are enabling more functionality with whisper models!

The file size difference comes from how HF implemented Whisper vs the initial code.
You can easily see here what weights are changing or embedding layer vs linear layer.

I am not 100% sure about if the size should be tripled, but there definitely should be a difference for sure.

Q: is there a way to reduce size to 2.9G for example? To be easier to download.
For this question, I am not really sure.

[doublex]

@DigitLib
Did you find a solution to reduce the size of the large-model to 3GB?

[doublex]

For example: This model (3GB) increases to 6GB:
https://huggingface.co/bofenghuang/whisper-large-v2-cv11-german

This is what I have tried:

from multiple_datasets.hub_default_utils import convert_hf_whisper
convert_hf_whisper('bofenghuang/whisper-large-v2-cv11-german', './large-de.pt')