How it works

Turning a recording into sheet music takes two different jobs: hearing which keys were pressed, and writing them down as readable notation. The first is now very good; the second is where most tools fall short.

1. Hearing the notes

A piano-only transcription model, Transkun, listens to the recording and outputs every key press with its start, end and loudness. In our tests it found 99–100% of the notes in synthesized piano recordings, and 106 of the 107 notes in the first 30 seconds of a studio recording of Bach's Prelude in C, with no extra notes. Basic Pitch, a general-purpose model used by free tools such as Loopaloo, found 103 of those notes but added 53 wrong ones.

2. Finding the beat

A beat-tracking model (beat_this) and the note onsets together place each beat, allowing the tempo to drift the way real playing does.

3. Writing the score

Our engraver chooses the time signature and where bar 1 starts, rounds each beat to readable rhythms (eighths, sixteenths, triplets), splits the notes between the two hands, gives sustained notes their own voice, and spells sharps and flats from the key.

4. You check and fix

The most common mistakes are a wrong time signature (for example 2/4 instead of 4/4), barlines shifted by a beat, or note values doubled. Each is fixed with one or two clicks — the score is rewritten in about a second, without listening again. On 8 synthesized test pieces we had not tuned on, at least 90% of the notes sat in the right place in the bar for 1 piece automatically, and for 7 of the 8 after the right one or two clicks. In the full piece you can also fix bar by bar where the player slowed down.

Getting the best result

• One piano only — no singing, drums or other instruments.
• A clear recording: close to the piano, little room echo, no talking.
• Steady tempo helps. Very free playing (long pauses, big ritardandos) is the hardest case for any tool.
• If the time signature is wrong, set it first; then move the barlines if bar 1 starts in the wrong place.

Transcribe a recording — any 30 seconds free.