On the music gen side of things, I have been keeping track of this guys (AceStep) progress, and it seems really exciting. They seem set on making a local/open-source Suno killer, and they have apparently have done a lot of good work on the architecture and dataset, hopefully the model indeed turns out amazing