Note that MIDI is a lot more effective when it comes to ML/AI, since it's multiple orders of magnitude less data. Daniel D. Johnson's (formerly known as Hexahedria, hired by Google Brain) model biaxial-rnn-music-composition is from 2015, requires very few resources for training or inference, and still delivers compelling, SOTA-or-close results wrt. improvising ("noodling") classical piano. https://github.com/danieldjohnson/biaxial-rnn-music-composit... You may also want to check out user kpister's recent port to Python 3.x and aesara: https://github.com/kpister/biaxial-rnn-music-composition (Hat tip: https://news.ycombinator.com/item?id=30328593 ).

Music generation from notation is pretty much the MINST toy-scale equivalent for sequence/language learning models, it's surprising that there's so little attention being paid to it despite how easy it to get started with.