We have constructed two different transformer models for music. The first one is only trained on Americana music trained on midi files scraped off the Internet. The second employs transfer learning from the pre-trained MAESTRO transformer. Below are links to music generated with the two models. The music files with same number have the same primer, i.e. the first three bars of music is the same.
This piece of generated music we find particularly interesting. We find that the predicted part of the music from the transfer model contains polypony, even though we only provide the music with a monophonic primer.