What I'm working on
I play alto sax and work on machine learning. With JazzML, I'm experimenting with models that generate solos using the Weimar Jazz Database. I choose the research direction and use AI agents to help write code and run tests.
Right now, I'm working on planning the rhythm and using chord charts to guide the notes. The solo and accompaniment also need to stay in time with each other.
The experiments
I've been working with MusicVAE-style models in PyTorch. I also have a simpler baseline that uses rhythms from the training data and fixed rules to add notes. That version runs, but it still hasn't met the rhythm and note-density checks.
The newer models are still experiments. I don't have a new version that has passed both evaluation and listening review yet.
How I check the results
I compare runs with different random seeds and keep the evaluation music separate from the training data. Each run has a record of the data and code it used so I can reproduce it.
I check the rhythms and how many notes the model generates, then listen to the solos. I've rejected runs that missed those checks even when the training loss went down.
Audio from an earlier version
These clips are from v6.8, an older version of the project. They aren't from the models I'm working on now.