What kind of audio works best?

A single melodic line, recorded reasonably cleanly. A played instrument — guitar, piano, synth lead — transcribes most accurately; singing and humming follow close behind. Whistling is the hardest: a whistle is nearly a pure tone, with little of the harmonic structure the model listens for, so it often needs the Eager setting. The model is polyphonic, so chords do work, but a dense full mix does not — there is too much overlapping content to separate. Drums and unpitched percussion have no notes to find.

Why did it miss notes, or invent them?

The sensitivity setting is the dial for exactly that. Shy only accepts confident notes and will miss quiet ones; Eager catches more and occasionally hears a note that was not played. Balanced sits between them and is where we landed after scoring the output against known-correct transcriptions. The shortest-note setting also drops brief blips, which are usually artifacts rather than notes.

Which model does it use?

Basic Pitch, released by Spotify under Apache-2.0. It is a small convolutional network — roughly 36,000 parameters in a 225 KB file — that outputs note onsets and sustains, which are then assembled into notes and written to a MIDI file. All of it happens in your browser.

Does it detect the tempo?

It does not detect tempo. The tempo you choose is written into the file so bar lines land somewhere sensible when you open it, but it does not change the notes or their timing — those are placed in real time, exactly where they were played.

Is my recording uploaded anywhere?

No. The audio is decoded and analysed in your browser on your own device. It is never uploaded, and nothing is kept once you close the tab.

Yes — completely free, no account and no limits. Everything runs on your device, so there is nothing for us to meter.

What is the piano sound?

Salamander Grand Piano V3 by Alexander Holm — a recorded Yamaha C5 concert grand, used under the Creative Commons Attribution 3.0 licence. Eleven sampled notes covering six octaves, about 380 KB in total, fetched once and then cached. If you would rather not download it, the synth voices work offline from the start. Either way the MIDI file carries the notes, not the sound.

Audio to MIDI — turn a played or sung line into notes

Drop in a recording of a melody and get the notes back as a MIDI file you can open in any music software. It works best on a single line played on an instrument — a guitar, a piano, a synth lead — and sung or hummed lines work too. Everything runs in your browser, on a model small enough that you will barely notice it downloading.

Your audio
Drop audio here — or click to choose
MP3WAVM4AFLACOGG
Your files are never uploaded — everything runs on your device.
Settings
MIDI

How to turn audio into MIDI

Add the audio

Drag in an MP3, WAV, M4A, FLAC or OGG, click to choose one — or press Record and play, sing or hum the line straight into your microphone. Nothing is uploaded either way.
Set the sensitivity

Balanced is the tuned default and suits most recordings. If quiet notes are missed — whistling especially — try Eager; if notes appear that you did not play, try Shy.
Download the MIDI

Press Play to hear the notes back before you commit — a wrong octave or an invented note is obvious in a second of audio. Then download a standard .mid file with the notes, timing and velocities, ready for a DAW or notation editor.

What you get

Best on a played line, and good on a sung one

A cleanly recorded instrument transcribes most accurately. Singing and humming work well too, and are often the fastest way to get an idea out of your head — they just need a steadier pitch to land as tidily. Whistling is the hardest case for the model, and usually wants Eager.

A real MIDI file, not an audio render

The output is a Standard MIDI File: individual notes with pitch, start, end and velocity. Open it in any DAW and every note is editable, transposable and re-voiceable.

Small enough that it just runs

The model is 225 KB — smaller than a typical web font. There is no lengthy download to warn you about before you start, and it works offline afterwards.

The model behind it

Powered by Basic Pitch, Spotify's Apache-2.0 note-detection model — a small convolutional network with about 36,000 parameters. It is polyphonic and instrument-agnostic, so it handles chords and any instrument. We tuned the detection thresholds against transcriptions scored on real recordings rather than shipping the stock defaults.

Hear it before you keep it

Press Play and the notes come back on a real recorded grand piano — Salamander Grand Piano by Alexander Holm, used under CC BY 3.0. About 380 KB, fetched once. Five synth voices are there too if you would rather not download anything.

Frequently asked questions

What kind of audio works best?: A single melodic line, recorded reasonably cleanly. A played instrument — guitar, piano, synth lead — transcribes most accurately; singing and humming follow close behind. Whistling is the hardest: a whistle is nearly a pure tone, with little of the harmonic structure the model listens for, so it often needs the Eager setting. The model is polyphonic, so chords do work, but a dense full mix does not — there is too much overlapping content to separate. Drums and unpitched percussion have no notes to find.
Why did it miss notes, or invent them?: The sensitivity setting is the dial for exactly that. Shy only accepts confident notes and will miss quiet ones; Eager catches more and occasionally hears a note that was not played. Balanced sits between them and is where we landed after scoring the output against known-correct transcriptions. The shortest-note setting also drops brief blips, which are usually artifacts rather than notes.
Which model does it use?: Basic Pitch, released by Spotify under Apache-2.0. It is a small convolutional network — roughly 36,000 parameters in a 225 KB file — that outputs note onsets and sustains, which are then assembled into notes and written to a MIDI file. All of it happens in your browser.
Does it detect the tempo?: It does not detect tempo. The tempo you choose is written into the file so bar lines land somewhere sensible when you open it, but it does not change the notes or their timing — those are placed in real time, exactly where they were played.
Is my recording uploaded anywhere?: No. The audio is decoded and analysed in your browser on your own device. It is never uploaded, and nothing is kept once you close the tab.
Is it free?: Yes — completely free, no account and no limits. Everything runs on your device, so there is nothing for us to meter.
What is the piano sound?: Salamander Grand Piano V3 by Alexander Holm — a recorded Yamaha C5 concert grand, used under the Creative Commons Attribution 3.0 licence. Eleven sampled notes covering six octaves, about 380 KB in total, fetched once and then cached. If you would rather not download it, the synth voices work offline from the start. Either way the MIDI file carries the notes, not the sound.

Working from a song rather than a single line? Split it first with the vocal remover, then transcribe one stem at a time.

Your audio

Settings

MIDI