For those of you who tried to read @groma’s paper of last year and struggled a bit, a more general post has been done here
The article is sadly not giving proper reference to this immense hyperactive field, as @groma kindly pointed to me, and was replied to along those lines by one of the world leaders here
I thought that might interest some of you.
p
Wow that unmix app is really surgical at getting the different components. I imagine that there is some serious training done on some big clusters. @groma do you know more?
you should check the website of the comparison - there are still a lot of spectral processing artefacts, but indeed for vocal there are good results (and @groma’s were good 
Hi, here’s a good resource on this topic
https://sigsep.github.io/
For big networks, you generally need a good GPU, as much memory as possible. Multi-GPU setups are not trivial, I don’t know if other people are using them. The network in the post is pretty big, but they mention using a single GPU.