How it works
From a recording to a score, honestly.
This is the whole method: what listens, what decides, what engraves — and where it gets things wrong. No accuracy percentage, because we have measured on too few pieces to hand you one in good conscience.
The first 30 seconds are free, in your browser, with no account.
The pipeline
Six steps, in the order your recording meets them. The first three happen entirely on your device.
01
Your browser decodes the recording
MP3, WAV or M4A is decoded by the browser's own audio decoder, on your device. The audio itself is never uploaded — not for the free preview, not for the paid transcription.
02
Mono, at the rate the model expects
The channels are downmixed to mono and resampled to 22,050 Hz. Leading silence is trimmed first — thirty seconds of stool-shuffling should not count as your free preview — and the excerpt is cut on a quiet moment rather than mid-chord.
03
Basic Pitch listens
Spotify's open-source Basic Pitch model finds the notes — pitch, onset, duration, velocity — in a Web Worker, so the page stays responsive. We run it at stricter thresholds than its defaults (0.7 onset, 0.4 frame), which we measured on solo piano: the defaults accept overtone ghosts that a piano generates constantly.
04
Notes become a score
Our own pipeline — written for this job, not part of the model — estimates the tempo from the onsets, the key from the pitch collection and the meter from the accents, splits the notes between left and right hands, and quantises everything onto a sixteenth-note grid with triplets, ties and rests.
05
Verovio engraves it
The quantised score is laid out on a grand staff by Verovio, the open-source engraver used by academic music projects — proper spacing, beaming and page breaks rather than a screenshot of a piano roll.
06
Three files come back
The same notes render as a printable PDF, a MIDI file for your DAW or piano app, and a MusicXML file that opens in MuseScore, Sibelius or Dorico. One transcription, three ways to use it.
What leaves your device: the notes, not the audio
The free preview is fully private by construction: your file is decoded and transcribed in two Web Workers on your own machine, and nothing is sent anywhere. You could transcribe the preview on a train with no signal and it would work the same.
The paid transcription runs the same way — the detection happens on your device too. What we receive is the list of detected notes, which our servers engrave into the PDF, MIDI and MusicXML. Your recording is never uploaded at all.
A transcription you do not buy is deleted after 24 hours; one you buy stays in your account for a year. The details are on the privacy page.
Estimated, or assumed
What it measures, what it guesses
Every transcription comes with a report, and every uncertain number in it is labelled — "estimated" when it was measured from your recording, "assumed" when it was not. The report never shows a confident figure it invented.
Tempo, when the playing allows it
The BPM is estimated from the onsets, with protection against double- and half-speed readings. When the playing is too free to measure, the report says "assumed" rather than showing a number we made up — a confident wrong tempo is worse than an honest guess.
Key and meter, with a confidence
The key signature comes from the pitch collection and the meter from accent patterns. Both carry a confidence in the report, and only 4/4, 3/4 and 6/8 are ever detected — anything else is written as the closest of the three.
Which hand played what
The left/right split around the middle of the keyboard is inferred, not heard. If it does not match how you play, moving notes between staves in a notation app takes seconds.
Where it fails
Automatic transcription is a first draft, not a finished score. These are the cases where the draft needs the most work — the free 30-second preview exists so you can see them on your own recording first.
Dense, fast polyphony
Many notes at once, moving quickly, is the hardest case there is. Virtuosic passages come out with gaps — this is the piece where proofreading earns its keep.
Heavy sustain pedal
Pedalled passages blur where notes end. On this model, sustained notes tend to come out too short rather than too long — check note lengths wherever you pedalled.
A lively room
Reverb smears note endings and can read as extra notes, especially in the bass. A phone on the music stand beats a good microphone across the room.
Anything that is not solo piano
Voice, drums or a second instrument share the piano's frequencies and produce wrong notes. The tool is built for one piano, alone.
Do not take our word for any of it: the examples page shows three real transcriptions, uncorrected — including the key signature it got wrong and the octave doublings it invented.
Why there is no accuracy percentage
We have measured this model, properly: against exact ground truth, across clean and degraded audio, on real piano pieces. That measurement is why the detection thresholds are what they are, and it is enough to say where the tool is strong and where it is not.
It is not enough to publish a number. The pieces in a test set are chosen, and yours is not — a percentage measured on a handful of recordings generalises to a stranger's living room about as well as a weather forecast for next March. We would rather tell you what was measured, on what, and let the free preview answer for your recording. The reasoning, at article length, is on the journal.
For the end-to-end walkthrough — preparing the audio, reading the report, proofreading the score — see how to convert a piano recording to sheet music.
Questions, answered
No. Decoding and note detection run in your browser, on your device — the audio never leaves it. For a paid transcription, only the detected notes are sent to us so the PDF, MIDI and MusicXML can be rendered.
The method, on your recording.
Thirty seconds, free, in your browser — the same pipeline the paid transcription uses, on the audio only you can judge.