https://gladia-research-group.github.io/multi-source-diffusion-models/

Read towards the middle for the audio separation part. The first part is about AI creating music, which is scary on its own.

If this gets released, it will make waves in the music industry and probably cause problems, but, it will also be amazing for creating samples from instruments deep inside an audio mix that you would normally not be able to get to. I'd love to try this out for that reason alone. Can't wait.

 

There's already multiple software solutions for audio separation, best one I found to date is demucs.
In really busy mixes it still is not accurate enough to my liking, but it does a scary good job every time a new version and/or model gets released.
From the examples on GRG website their separation seems a bit basic, I guess it could have big issues with busy mixes.

 

I've noticed that most demos of AI music separation tools use source music with a lot of space between the instruments. I'll be impressed when they can do a decent job separating the bass and guitar with a Black Sabbath song.

Not criticising, they're getting better by the day.

(comment generated by CrinklebumpsGPT)

 

This is already available in the market(Google "AI stems") and I'm not understanding how this is scary. This is evolution and comparing our own stems to industry standard mastered stems should give us better, more accurate tools and thus, results.

 

You're missing the satire in my "scary" statement, but ok.