A melody no one has ever heard.

Twelve melody models so far, from a small LSTM that learned 24 R&B tracks to a 96 million parameter Transformer trained on 117,682 melodies. Pick one, set the dials, and it writes a tune nobody has heard before.

How it works →
§ I

The composer's desk

Pick a voice, set the dials, press compose. The newest model reads six words before it writes a note (key, register, motion, density, repetition and mode) and does its best to honour them. Leave a dial on Any and it makes its own mind up. You get the score and the audio.

The desk is clear. Compose something and the score will arrive here.

§ II

The songbook

Everything anyone composes here lands in the songbook, filed by opus, with the newest model first. Pick an opus to hear how the models changed, from the first little LSTM to the one on the desk now. Each piece is a one off. Put the dials back exactly where they were and you still won't get it twice.

§ III

How it got here

OPP. 2–5

The early ones

LSTMs, one per genre, each picking up note-to-note habits from a few hundred tracks. Charming, repetitive, and every so often completely lost.

OPP. 9–11

The genre years

Tens of thousands of real melodies with a genre word on the front of each one, then op. 11 at 96 million parameters on a cleaned-up corpus of 91,000. The genre word moved the numbers. Nobody could hear it.

OP. 12 · NOW PLAYING

The rebirth

Genre is gone. Six dials measured from the music itself take its place, and you can hear every one of them. Trained on 117,000 melodies, 26,000 of them folk and jazz tunes. Read how it works or follow the journal.