I tried to generate something that would have more harmonic form and less individual staccato notes. The longer samples are more or less addressless meandering, though. The shorter ones appear more intact, but in fact, they are just snippets from a longer, less coherent sequence.
On the other hand, apparently TensorFlow are soon releasing something called Music Transformer, which I’m sure will blow my attempts out of the water, as it is built with long-spanning structure specifically in mind. It’s open source and it’s very cool.
