> but neural-networks can separate them in practice

That's a massive stretch.

They are able separate them sort of, but they are nowhere close the original quality of the individual tracks.