Sort of outside the topic, but I’ve done a lot of 15ips/7.5ips conversions over the years using Elastique. Varying settings of formant.
Not necessarily in the neighborhood of difficult…but I’ve felt for some time that a good code-training track would be the mono 45 of the Beach Boys’ “Let’s Go Away For A While”.
Just think, a menu button that shows-
unmix bass
unmix upright bass
unmix rhythm guitar
unmix vibes
unmix saxes
unmix violins
unmix violas
unmix bass drum
unmix toms
unmix coke bottle guitar
unmix oboe
unmix timpani
unmix snare
unmix slap stick
unmix flutes
Beats doing everything one at a time spectrally
People do create their own models. You simply need access to a lot of stemmed out audio and some decent GPU resources available to you to carry out the training but not very difficult when you get into it. This was how some of the recent Peter Jackson demix models were created for the Beatles Get Back and Revolver demix process.
Thanks for your detailed exposition; I had not appreciated the extent to which the AI training might impact the unmixing process. That also may explain why SL 9 seemed to handle the bass extraction from this track better than SL10, although SL10 did significantly better with the guitar layer.