Extract stems using same song as training data?

This is a long shot, but I wonder if there is any way of extracting stems from songs using spleeter without having access to hundreds/thousands of examples of training data.

I have tried to use phase cancellation from different parts of the same song to isolate individual elements but the results are not spectacular. Even for electronic music there seem to be slight phase offsets which always punch through. In some cases the phase inversion technique is ok but not enough that it an be a viable way of extracting stems.

I have had success using spleeter in isolating drums, vocal, bass, piano and other but sometimes I need more than 5 stems. My goal is to isolate as many elements as possible such that I can recreate tracks in my DAW and get insight on the underlying song structure.

Perhaps there is a way of combining phase inversion, onset detection and machine learning/NN to extract features from a song if a large number of training data is not available? Perhaps parts of the same song can be used as examples of training data?

Thanks.