Cut-Up

BPM

Time /

Swing

🎛 Mix

The desk

Five channels — the tune, the voice on it, the chords, the bass and the kit — each with a fader, a pan, a three-band EQ, a squeeze and a send to one shared plate. Everything moves while the song is playing, and the mix travels with the song: an exported file comes out the way you left it here.

Master Reverb size Reverb damping

Desk in circuit

The transport's part switches above leave a part out of the arrangement altogether; muting it here only turns it down. Soloing is the quickest way to hear what one instrument is actually doing.

🫀 Feel

Humanize

Loose ↔ tight Nobody plays on the grid. This pulls every note a little off its step — and lets each player be off it in their own way: a drummer who rushes the hats, a bass that sits behind the beat. At zero it is a machine again.

Feel seed

The deviations are seeded like everything else here, so the same seed is the same take — and the MIDI export is the take you just heard rather than a second, differently sloppy one.

🎼 Tuning

Temperament

System Divisions of the octave Twelve is the piano. Anything else keeps the twelve notes of the key and the chords you know, and puts them on the nearest steps of a finer ladder — 19 and 31 flatten the thirds, 17 sharpens them, 24 lands back on the piano. The octave and the tonic never move.

Out of tune

Instrument detune Note-to-note drift

Tuning seed

Detune is a fixed offset per instrument — a section that tuned five minutes ago. Drift is per note, which is a singer rather than a piano. Both come off the seed, so a piece keeps the same slightly-wrong tuning every time you play it.

Teach it what you like

It deals two or three sections out of the composer's own deck. Listen, say what you think, and it deals again. Nothing here is a quiz with right answers — it is only ever learning your opinion, and the only way it can be wrong is by disagreeing with you.

1–5 answers the highlighted question and moves on, Space plays whatever that question is about, Enter saves the hand and deals the next. Leave anything you have no view on unanswered — a shrug is not a score, and it is better for the model to hear nothing than to hear a guess.

This is the thorough way and it is also work. The quick way is the 👍 and 👎 on every section card and every song, on the Song and Library tabs: one click, on music you were already listening to, saying only yes or no. Those go in the same file and count for less — a thumb nudges the model, a rated hand moves it — so you can teach it entirely by working and never open this tab at all.

The model

How much say it gets when composing Scaled by how well the model did on rounds it was never trained on, so one that has not beaten a coin has no say at whatever this is set to. Run npm run taste instead of npm start and every answer is written straight to data/taste.jsonl; training then rewrites models/taste.json, and both are yours to commit.

Title ideas

Keepers

Melody

Sections

Arrangement

Saved songs

Saved sections

The desk

Humanize

Temperament

Out of tune

On the table

Roll it again

What to roll

Key

Count

Variations

Seed

Rename

Are you sure?

Export audio