This is an AI/ML system by Meta which can separate recorded music into four “stems” - drums, bass, vocals and other. It’s not the first one (Deezer open sourced Spleeter a few years back which is now widely used) but I think the results from Demucs are super impressive - it sounds much “crisper” than previous systems. Kind of blows my mind that this is possible!
On my M1 Max it runs at 5-10x real-time if you pass a “-j 10” argument (and makes it real hot, lol) so it’s fast enough to be usable. You can also try it out online e.g. here: https://huggingface.co/spaces/akhaliq/demucs
I thought I’d share this as in the Kate Bush thread someone was talking about being able to get acapellas for tracks.
On my M1 Max it runs at 5-10x real-time if you pass a “-j 10” argument (and makes it real hot, lol) so it’s fast enough to be usable. You can also try it out online e.g. here: https://huggingface.co/spaces/akhaliq/demucs
I thought I’d share this as in the Kate Bush thread someone was talking about being able to get acapellas for tracks.