Intuitively (to me), it feels like a clever variation of something like Needleman-Wunsch could pull this off without using machine learning (as long as you have the right encodings). Can someone enlighten me?
There is only an embedding or re-presentation in some less informative form. ie., you have to choose some partial measurement system and apply this to the images of scores to produce a numeric representation.
This is, in part, why a midi-file is always going to sound worse than an orchestra operating from sheet music. The musicians understand the intensions of the composer.