Concept: Using parts of the track without vocals as reference for getting full instrumental

[Chikko5]

The track has segments where there are vocals and where there aren't. Algorithm analyzes parts of this specific track which don't have vocals and based on these, removes vocals in other parts of the track.
Is this even a thing?

[itadaku23]

That's a pretty nifty idea, but that's not how these models, or Fourier transformers in general work. The models don't get trained on separate instruments rather they are a probabilistic state sinusoidal pattern matching and mathematical extraction machine. That's why you can't get piano / strings / other similar-range timbre instruments to separate. That's why it's always drums / bass / mid range / vocals as stems. Music played by instruments of similar 'range' cannot be separated any more than what is. Not much you can do about it.