How audio becomes sheet music, and what to check

The sheet-music tools listen to a recording, write down the notes, and lay them out as a score, all inside your browser. Finding the notes is the part that works well; fitting freely timed playing into bars is the part that needs your help. This guide shows what the engines measured and how to get a draft that needs less fixing.

What the engines measured

Before release, both engines were run on five piano parts from Schumann’s Dichterliebe rendered from CC0 OpenScore scores, and on the same audio with simulated room reverberation and background noise added. The piano engine is Transkun V2 compressed to int8; the other-instruments engine is Spotify’s Basic Pitch.

Note-onset F1 on the 2026-09-30 engine-selection benchmark, Python reference run (1.0 is perfect)
Engine	Clean piano	Simulated room	Model size
Piano (Transkun V2)	0.993	0.988	24.3 MiB
Other instruments (Basic Pitch)	0.874	0.760	0.2 MiB
Basic Pitch on a bowed melody	0.432	—	0.2 MiB

Synthesized renders and simulated reverberation are kinder than a real piano in a real room, and no real room recording has been measured yet, so treat these numbers as an upper bound.

Standard and High accuracy

High accuracy does not change the notes: they are still the piano engine’s. It has Basic Pitch listen as well and marks in red every note Basic Pitch did not confirm. On the benchmark, 11% of the notes turned red on the clean recordings and 26% with simulated room sound, and only 5–6% of the red notes were wrong; but 88–98% of the piano engine’s wrong notes were red, and at most 0.1% of the black notes were wrong. So red is where to look first.

Standard: the piano engine alone.
High accuracy: the piano engine plus Basic Pitch, about 1–3 seconds more for a three-minute piece.
The red is written into the MusicXML, so it stays red in MuseScore; a switch in the score options hides it.

Getting a draft that needs less fixing

Record to the metronome with the one-bar count-in: the tempo and the first downbeat then come from the metronome, corrected for the audio latency the browser reports, instead of being estimated.
Wear earphones, or turn the click sound off and follow the visual beat; a click that reaches the microphone is transcribed as notes.
If the tempo warning appears, type the tempo; the score redraws without transcribing again.
Pick the coarsest note grid the music allows; a sixteenth grid on an eighth-note piece invents rhythms.
Rubato playing quantizes poorly; the MIDI keeps your real timing even when the bars do not.

Editing and coming back

The score is a draft. Download the MusicXML, fix it in a notation editor such as MuseScore, and open the edited file with MusicXML to PDF to print or transpose it. A keyboard performance saved as MIDI goes straight to MIDI to Sheet Music with no transcription at all.

Back to PrismMesh