It makes me wonder whether the music generation should be stratified -- a coarse model lays out where parts like verse and chorus are, what distinguishes them, how to transition, etc., and then a finer-grained model fills in the details.