Diving into audio processing AI techniques as a newbie

[julvst]

Hi!
I'm a software engineer, with a bit of experience in DSP. I find this field very interesting and I'm looking forward to learn more about it.
My experience in DSP comes from being involved in the development of hearing-related algorithms but always as the programmer who implemented/maintained the models from my DSP colleagues. However this allowed me to learn about the topic, so I'm fairly familiar with time/spectral domain processing, and can understand some of the underlying maths.
At the moment I'm doing some research about drum separation, looking at what tools are already there. I'd like to study them and try to reproduce some of their processing.
Recently I've been playing around with Acon Digital's Remix:Drums and I'm impressed by the results I get from it so I'd like to understand more about it. In their marketing they claim it is AI-powered, and I'm very curious to learn about this. Can anybody point to any literature/blogs where I can study AI drum separation techniques? I'm also interested on any information about classic DSP methods for drum separation or anything you consider is useful to dive into this topic.
Thanks

[quikquak]

I’m sorry for the cop out answer, but have you just tried google with “where I can study AI drum separation techniques?” In the search bar? People don’t seem to discuss AI much on here.

[julvst]

AUTO-ADMIN: Non-MP3, WAV, OGG, SoundCloud, YouTube, Vimeo, Twitter and Facebook links in this post have been protected automatically. Once the member reaches 5 posts the links will function as normal.

I’m sorry for the cop out answer, but have you just tried google with “where I can study AI drum separation techniques?” In the search bar? People don’t seem to discuss AI much on here.

Yes I did that but all I could find is the user / music production perspective. That's why I thought someone in a forum like this one could hopefully give some hints.
The only interesting thing I could find during my research was this paper: https://www.sciencedirect.com/science/a ... via%3Dihub (https://www.sciencedirect.com/science/article/pii/S0167865524001351?via%3Dihub), but I don't have the basic knowledge about the topic to understand it...

[Marvinh]

I don’t know too much about source separation I tried when I was in school but very hard to understand at the time.

They called it blind source separation.

What I would try now. try to identify transients and then identify what type of percussion it is . Then play it in a different order.

[julvst]

What I would try now. try to identify transients and then identify what type of percussion it is . Then play it in a different order.

I've already done some experiments like this by doing some spectral analysis. I found that the amplitude spectrum of just a few samples of the transients are enough to differentiate a drum type. First I store the spectrums of a drum kit, then I compare these to the transients of my real time input. Works for a very dumb drum machine that doesn't have much variation in the sounds and that can only play 1 sound at a time

[quikquak]

Yeah, doing it analytically would mean finding the harmonics of all the simultaneous transient tracts as well. For drums that’s not easy, as there’s a lot of high frequency noise or random transients. You may end up just removing other instruments from the track one at a time, and there will be a lot of ‘bleed’ with mistaken transients (from guitar plucks etc).
With AI, it’s all about your training data, giving it a lot of examples of music with and without its percussion. You may end up needing a combination of techniques. Spleeter has pre trained data:
https://github.com/deezer/spleeter
And this one might be worth looking into:
https://github.com/adefossez/demucs
I haven’t tried training my own AI yet, but input data is king! You may need a hundred thousand example tunes.
Those examples will have to be high quality stems, so it’s not easy to find hours and hours of that kind of data.
(Edit) you might also like this link to a some data, probably not enough, but you never know…
https://sigsep.github.io/datasets/musdb ... ssed-stems
("not be used for any commercial purpose without the express permission of the copyright holders")