My first music generation app

After 3 years of training music generation models and interacting with them through CLI, this is my first attempt to build a user interface for them. Playing with it feels really great, it feels like the model is actually usable.

https://github.com/eri24816/midi-gen

The user interface has a piano roll where you can make music. When needed, you can ask AI to generate a bar for you. It will offer multiple options for your choice, which can be further edited. Occasionally, there emerges music that sounds amazingly good but you didn’t think of. I believe that kind of “accident” is the main value of this AI in music composition.

In this interactive scenario, I view the AI merely a source of inspiration while the human remains the primary composer. The AI has yet no ability to percept the global structure (I’m currently working on a new model to address that). You have to fix glitches in the generation, and sometimes you need to specify chords and velocities for each bar to increase the chance of a reasonable outcome. You can also try asking it to generate all bars without any hints, and it will generate something locally good but globally weird.