Text to MIDI

Published October 2026

Text-to-MIDI tools usually return one clip: a chord progression or a melody. A song needs more than that: a form, and parts that know where they are in it, so the bass rests in the intro, the drums fill into the chorus and the lead waits for the verse. Ask the studio chat for a song and the arrangement comes back as MIDI, one track per part, across the whole form, with each part also rendered as audio, in a session you can export to your DAW or as MIDI files.

What "a 60-second trap song" becomes

This is the arrangement the song engine wrote for that request on 29 October 2026, in 0.48 seconds:

style   trap / drill      tempo 145 bpm (the pack's range is 130–160)
key     B minor           36 bars, 59.6 s
form    intro 4 · verse 12 · hook 8 · hook 8 · outro 4
chords  F# | F# | Bm | Bm | Bm BmMaj7/A# | Bm7/A Bm6/G# | C | C | …
parts   Drums    803 hits    in every section, fills into bars 5, 17, 25, 33
        Bass      86 notes   enters with the verse
        Rhodes    64 chords  plays from bar 1
        Synth     92 notes   the lead, from the verse

The move from B minor to C is the pack's Phrygian shuttle (i to ♭II, a drill staple), and the chromatic line under the tonic in the verse's first bars (B, A♯, A, G♯) is a line cliché, one of the pack's occasional devices. The engine records how it decided each thing (the pack, the form, the tempo range, the key, the fills), so a result can be explained rather than taken on trust.

A request with a key, a tempo or a band ("in D minor at 92 bpm with Rhodes and upright bass") is written to them. Without a length you get about two minutes.

Rules, not guesses

The notes do not come from a language model predicting MIDI tokens. They come from an arrangement engine with 45 genre packs, each describing its genre's forms, harmonic rhythm, chord habits, grooves, bass styles, comping and lead feel. The chat model's job is to read your request: the genre, the length, the band, whether it is sung. That split is why the output is always valid MIDI in the key you asked for, why the parts agree about the chords, and why the same request with the same seed gives the same song.

The country song on the overview page came from the same engine: 20 bars in F major at 110 bpm, with 321 drum hits, 41 bass notes and 284 guitar notes, every one of which arrived in the Logic and Ableton projects as MIDI.

Getting the MIDI out

• As a DAW project. Export / Share → Logic Pro, Ableton Live, Pro Tools, FL Studio or REAPER. Each part arrives as a MIDI track next to its rendered audio track.
• As MIDI files. Download MIDI in the same panel saves one track or the whole song as a .mid, ready to drag into any DAW or notation program.
• Just the notes, from an agent. An AI agent connected over MCP can write the parts itself and build the project from them with doseedo_session_from_stems, which takes MIDI tracks alongside the audio. To check the harmony of notes it wrote, at no cost, it calls doseedo_analyze_midi.

Text to MIDI, then to sound

MIDI on its own is silent until something plays it. Each part is also rendered by an instrument model trained on real playing, rather than a General MIDI SoundFont, and the render is checked against the notes: the receipt says how many written notes it heard. To render a MIDI file you already have, use MIDI to audio; the instruments are listed on the Sources page.

Related

• Songs straight into a DAW session
• MIDI to audio with real instruments
• Render MIDI from an AI agent
• The other direction: audio to MIDI