jarredou
commented
1 month ago Hi, and... Wow, I wouldn't thought that this small experimentation could lead to this ! Amazing :)
I think it can be even better as this was kinda only a test run, training could have been pushed further leading to a bit better quality, and the training configuration was really lightweighted. Also, the dataset I've created for this model had some issues. I'm currently working on a new larger drums separation dataset without these issues.
With that better dataset and a config more focused on quality than on being lightweighted, we'll hopefully have some better model(s) in the coming months.
I'd be really glad to help you as much as I can even if I'm not a scientist, you can contact me at jm.jarredou@gmail.com
xavriley
Hi - thanks so much for releasing the 6-stem drum separation model. I've been looking at drum transcription (audio-to-midi) as part of my PhD research and after playing around with this model for a couple of days I have something that looks like SOTA on the MDB drum dataset! (82.5% F1 - a 1.5% improvement on previous SOTA)
Would you be interested in working on a paper together to describe the details? I've written some postprocessing code to extract the onsets and clean things up a little but I feel like you and @aufr33 deserve the credit for its success. I'm happy to take care of writing up the paper and doing evaluations.