[Discussion] training from real 4/8 track recordings or real "stems"

I have a fairly extensive library of stems and 4 and 8 track samples of songs. Put simply, how do i train this with the new data to supplement what is already there?

also, this is literally the most novel thing i've seen in the sound engineering / hacking space since binaural beats with cool edit pro 2.0, probably since i first heard midi with soundfonts, too.

I hope that you don't get paid out before this gets a lot of love, development-wise. I reckon karaoke and record labels will be all over this technology before any of us can say boo :-D