DNN-based_source_separation icon indicating copy to clipboard operation
DNN-based_source_separation copied to clipboard

Published 20 hours ago • verified publisher icon tky823

Join efforts?

Open mpariente opened this issue 4 years ago • 4 comments
trafficstars

Hi @tky823, nice repo !

We'd welcome most of this code in Asteroid if you'd like to contribute :smiley: Would you?

Cheers,

mpariente avatar Jan 25 '21 18:01 mpariente

Hi @mpariente, I'm interested in DNN-based source separation. I would be happy to contribute to Asteroid. What is the first step?

tky823 avatar Jan 26 '21 12:01 tky823

Great!

What is the first step?

Well, there are architectures here that we don't have in Asteroid, it would be super cool to have them there. For example, DANet and ADANet, or the Furca architectures.

Have you had a look at our source code and the organization? There are some constraints for the code to be consistent between models (easier maintenance and lower barrier for users). Anything would be welcome actually!

mpariente avatar Jan 26 '21 13:01 mpariente

I have not read the constraints, so I will have a look.

I've confirmed that DANet separates the mixtures. But I have not checked the score (e.g. SDR improvement). As for ADANet and the Furca, very little progress has been made in their implementation, sorry.

I use a free Google Colaboratory plan to make sure my implementation works, so it may take some time, but I'd like to join efforts.

tky823 avatar Jan 26 '21 16:01 tky823

There are also some Asteroid users that might be able to run your architectures on bigger hardware.

Feel free to join Asteroid's slack workspace to start discussing :wink:

mpariente avatar Jan 26 '21 18:01 mpariente