AI music transcription
This is not speech-to-text. Music transcription means getting the notes back out of a recording — pitches, rhythms, bar lines, chords — so you can read the part, edit it, and play it again.
What automatic music transcription actually is
Automatic music transcription — AMT, in the literature — is the problem of recovering symbolic music from audio. The output is not a document of words; it is note events with pitch, timing and duration, which then have to be written down as readable notation or usable MIDI.
Those are two genuinely different problems, and most tools only solve the first. Detecting that a pitch happened is one thing. Deciding it is a dotted eighth starting on the and-of-two in bar nine, on a stave in E flat major, is another.
One model does not fit every instrument
A grand piano and a tenor saxophone fail in opposite directions. Piano is polyphonic with sharp onsets and heavy sustain-pedal overlap. A saxophone is monophonic but full of the things pitch trackers hate: scoops, bends, subtones, growl.
A single general-purpose transcriber has to compromise for both and does neither well. So the take is routed to the model that suits it — a dedicated polyphonic model for piano, register-conditioned pitch tracking for solo winds, brass, strings and bass, and a general polyphonic model for guitar and ensemble material.
If you know what you recorded, say so and the routing follows. If you do not, it is inferred from the audio.
Tempo is part of the transcription
Notes without a grid are not a transcription, they are a list. The tempo, downbeat and meter are detected first and the notes are fitted against that — and the tempo map is allowed to move, so a performance that slows into a cadence transcribes at the tempo it was actually played, instead of being flattened onto a click.
One part, or the whole band
Give it a single stem and you get that part. Give it a full mix and the sensible route is to separate first — vocals, drums, bass, guitar, keys — then transcribe each stem on its own. Parts that overlap in a mix stop overlapping once they are separated, which is why this is the biggest single accuracy win available.
The transcribed parts are then engraved together as one score: bracketed system, aligned bar lines, the right clef for each part, a real key signature, and transposing instruments written the way their players read them.
What you get out
Notes, as an editable multitrack MIDI arrangement you can correct by hand. Sheet music, as a multi-stave score with clefs, key signatures and written transposition. MIDI per part, reflecting your edits rather than the first guess. And a Logic or Ableton session with the stems on named tracks, the tempo map and a chord guide.
Transcribe something
Upload a recording and get the notes back. Free runs daily, no install.
Open the transcriber →FAQ
What is AI music transcription?
AI music transcription — also called automatic music transcription, or AMT — converts recorded audio into symbolic music: notes, rhythms, bar lines and chords. It is not speech-to-text. The output is MIDI and notation you can edit and play, not a text document.
How accurate is automatic music transcription?
Accuracy depends far more on the material than on the model. A clean solo piano or a single-line horn recording transcribes very well. A dense mix, heavy distortion or a lot of reverb is harder, because the notes genuinely overlap in the signal. Separating the mix into stems first and transcribing each part on its own is the single biggest accuracy win, and every result stays editable.
Can it transcribe a whole song with several instruments?
Yes. The mix is separated into stems, each stem is transcribed on its own, and the parts are engraved together as a multitrack score with aligned bar lines, per-part clefs and correct written transposition.
Do I need to install anything?
No. It runs in the browser — upload a file and get notes, MIDI and a score back. There are free runs every day with an account.
What formats can I export?
MIDI for every transcribed part, plus a full Logic or Ableton session with the stems on named tracks, per-part MIDI, the detected tempo map and a chord guide.