Improved FullSubNet Usage

[omarthen3rd]

Running downsampled 48khz data on the existing model/inference doesn't seem to be a problem. It works really well!

However with the new improved model, I have a suspicion that I'll need to retrain data at a higher rate to use it since I get a lot of issues if I try to throw it in as is (even with updated [model.args] in a new inference config file)

My question: is it possible to use the new model with the existing datasets?

Cheers!

[haoxiangsnr]

Hi, @omarthen3rd. Thanks for your attention.

I did not save the pre-trained new model with 24KHz and 48KHz data. If you want to train it yourself, the necessary step is to modify [model.args]. How to modify the specific values? Check the comment for suggestions.

[omarthen3rd]

Ah I figured I'd have to retrain it, I kept getting shape mismatch errors after attempting to make my own inference.toml for 48khz audio based on those comments.

I'll simply convert the existing dataset to 48khz and retrain with new parameters. Thanks for your help, will keep you posted!