A melody no one has ever heard.
Neural sequence models, from early LSTMs to a REMI-tokenized Transformer, each composing single-voice melodies that are engraved here as they arrive.
How the instrument works →The composer's desk
Choose a model and a voice, set the temperament, and press compose. The score arrives with its audio.
The desk is clear. Compose something and the score will arrive here.
The songbook
Recent compositions, kept as a public register. Each entry is one of a kind: no setting of the dials will ever produce it twice.
Notes on the method
The early studies
Long short-term memory networks, one per genre, each learning note-to-note habits from a small corpus. Charming, repetitive, occasionally lost.
The conditioned studies
Trained on tens of thousands of true melodies with a style token opening every phrase: the genre dropdown genuinely steers these models, and both serve here as research previews.
The eleventh
Scale and clean: roughly 109,000 melodies, quality-filtered by the same metrics the output is judged with. Read the notes on the method or follow the journal.