Audio to notes: how to convert a song into note names

Converting audio to notes takes a recording, which is just a stream of sound pressure measurements, and turns it into a list of musical events: this note started here and ended there. It sounds simple, but it’s one of the harder problems in music technology. Knowing how it works helps you get better results and understand the mistakes.

How audio-to-notes conversion works

A sung or played note isn’t a single frequency. A guitar A2 vibrates at about 110 Hz, but it also produces quieter overtones at 220 Hz, 330 Hz, 440 Hz and so on. That’s what makes it sound like a guitar and not a flute. When several notes sound at once, all their overtones overlap. The converter’s job is to work out which combination of notes would produce the mix of frequencies it hears.

Noteform does this in a few stages, all inside your browser:

Decoding. The browser decodes your MP3, WAV, FLAC, M4A, OGG or WebM file into raw audio and mixes it to mono.
Pitch detection. In Polyphonic mode, Spotify’s Basic Pitch neural network estimates, for every moment and every piano key from A0 to C8, how likely it is that the note is sounding and whether it has just started. In Single melody mode, the pYIN algorithm tracks the single strongest pitch over time instead.
Note building. Those moment-by-moment estimates are joined into notes with a start and an end. Very short fragments are dropped, and repeated attacks of the same pitch are split into separate notes.
Cleanup. Notes that look like overtones of a lower note are flagged overtone?, and notes with weak evidence are marked ?, so you know what to check.
Chords. In Polyphonic mode, the detected notes are grouped over time and matched against chord shapes to estimate chord names.

What the result looks like

Each note comes out as a row with its start time, end time, length and name, for example 00:03.250 00:03.710 0.460 E4. Times are measured from the start of the recording in minutes, seconds and milliseconds. The name uses scientific pitch notation, where middle C is C4. The downloadable text file adds a header that records which part of the file was analysed, which mode was used and the tuning assumed, so you can always tell how a transcription was made.

The output is not sheet music. There are no bar lines, time signature or note values like quarter notes, because those depend on tempo and musical interpretation that the recording alone doesn’t settle. The times are exact performance times, which makes the list ideal for learning by ear and for feeding into other tools.

Why conversion makes mistakes

Overtones. A loud low note’s harmonics can look like real notes an octave or a twelfth higher. This is the most common error.
Dense mixes. When many instruments share the same range, notes mask each other. Quiet inner parts may be missed entirely.
Drums and noise. Percussion has no clear pitch but can trigger brief false notes.
Expressive pitch. Vibrato, slides and bends move between semitones, and the converter has to pick the nearest note.
Tuning. Noteform assumes standard tuning, A4 = 440 Hz. Recordings far from that can sit between two note names.

Any tool that claims perfect audio-to-notes conversion for every recording is overselling. The realistic goal is a draft that’s right most of the time and clearly marked where it might not be.

How to get the most accurate notes

Start from the cleanest audio. A lossless or high-bitrate file beats a low-quality stream. Isolated instruments convert far better than full mixes.
Pick the right mode. Single melody for one line, Polyphonic for chords and multiple parts.
Convert short sections. Select a phrase or a verse rather than the whole track.
Review the flagged notes first. Remove the flagged overtones in one click if they’re wrong, then listen to the notes marked ?.
Fix and recompute. Correct notes, add missed ones with + Add note, then press Recompute chords.

Converting audio to notes privately

Many online converters upload your file to a server. Noteform doesn’t: the decoding, the neural network and the cleanup all run in your browser, so your recording never leaves your device. It also means conversion speed depends on your computer: on a recent computer the analysis typically runs several times faster than the music plays, while phones and older machines take longer.

Have a piano recording? Read how to transcribe piano music online for piano-specific tips.