For now, between the state of the art source separation models (e.g. demucs) and transcription models (e.g. Magenta's MT3) the last mile seems to be MIDI -> Mus

For now, between the state of the art source separation models (e.g. demucs) and transcription models (e.g. Magenta's MT3) the last mile seems to be MIDI -> MusicXML IMO. But yes, I suspect it'll become more end-to-end ML in time.