Cool - but there is no proof that these demo songs are out-of-sample, and the trained model does not seem to be available in the github.
If I am understanding correctly, the github code doesn't contain the code to train the model? [1]
The paper also doesn't seem to be peer-reviewed? (I guess it is pre-review, since it is on arxiv only 2m ago)
No disrespect to the authors - looks like a really cool project - I am just a bit suspicious without seeing any performance on user-submitted data.