For the generation part, it might be a good idea to separate "music" into its parts (like bass, drums, melody, chord-rhythm, chord-ambient-background, and other

For the generation part, it might be a good idea to separate "music" into its parts (like bass, drums, melody, chord-rhythm, chord-ambient-background, and other stuff like secondary melodies, horn section arrangements or whatever). A unique model for each one of those could learn differently. Each of those parts could be broken down even further too.