search

facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation
MIT License
8.16k stars 1.03k forks source link

SDX 2023 Challenge #471

Open Dyslexicon opened 1 year ago

Dyslexicon commented 1 year ago

I see the competition ends May 1st; should we be expecting some new state-of-the-art separation models for Demucs soon? Any idea if these would be released immediately or how far out are we talking?

adefossez commented 1 year ago

I didn't take part in the competition, so no release on my end, but it is possible some candidates will have used Demucs!

Dyslexicon commented 1 year ago

So, The 1st and 2nd place winners have not/will not share their models. 3rd place model MVSep-MDX2023 is a massive improvement over htdemucs_ft in drum/bass/other stems but the vocal stem is trained on lossy which causes spectral truncation and generates garbage frequencies, rendering the entire model effectively useless as-is. This model needs to be rebuilt using full-spectrum vocal training data, at which point it would be the world's best public model.

As of right now we have gained nothing from the SDX 2023 challenge, quite disappointing. IMO this is a call for Demucs to leapfrog the hoarded and flawed advancements of this year. I would be most pleased to see Demucs v5.0 (or simply some new 2023 Demucs models) trounce the SDR results of this year's top SDX-challenge models. The "other" stem needs work, lagging behind the rest of the stems in accuracy. Lots of training on Rock music would likely improve the accuracy of this stem.