Exploration
Three model families
LSTM, GRU, and causal Transformer experiments on a reproducible MIDI pipeline.
3architectures explored
MIDI continuation
A melody in.
A new direction out.
Pick a short musical idea below. AudioGroove will use it as the beginning of a new MIDI continuation.
Start with one of these
curated from the project datasetHave your own?
Behind AudioGroove
A research project you can actually listen to.
I built AudioGroove to explore how sequence models learn musical structure from MIDI, then shaped the trained model into this small interactive experience.
The interface is intentionally simple. The training work lives underneath it, not in the way of the music.
Training
Scaled GRU experiments
Successful training runs at 500, 1,000, and 2,500 songs on HPC infrastructure.
2,500songs in the largest trained run
Result
GRU-large model
The largest preserved scaling result, with its checkpoint, vocabulary, and configuration kept together.
6.1306best validation loss