Meh. There’s plenty of deep learning music generation stuff out there, this is still a really cool approach I think it would be cool to combine the two. Instea

Meh. There’s plenty of deep learning music generation stuff out there, this is still a really cool approach I think it would be cool to combine the two. Instead of generating raw midi, your GAN or reinforcement learning agent or whatever could try to generate sequences of transformations to melodic fragments. Neural program synthesis type stuff. Or maybe one could build an automatic music analysis tool that can start from the score and try to infer the program that generated them. (Is that a thing already?)