At the NeurIPS workshop of Audio Imagination, we present a supervised method as a preliminary step towards answering these questions.
openreview.net/pdf?id=Zhc6Z...

How to train a model to extract separate entities of representation associated to individual sources of a music mixture?

How to also divide each entity into subspaces of pitch and timbre?

How to then have the model take arbitrary combinations of these building blocks to sample novel mixtures?