Audio to MIDI

What it can and cannot hear

This follows one pitch at a time. Give it a hummed melody, a bassline, a sung phrase, a flute or a monophonic synth lead and it will track the fundamental accurately. Give it a chord, a full mix or two instruments at once and it will return something confident and wrong, because there is no single fundamental to find.

That is not a rough edge to be fixed later. Polyphonic transcription — pulling four simultaneous notes out of a piano chord — is an open research problem, and the tools that attempt it are large neural networks with mixed results. A page that pretended otherwise would waste your time more expensively than one that says so.

If what you have is a full track and you want one part out of it, run it through the stem splitter first, then bring the isolated part back here.

How the pitch is found

The detector is YIN, which is autocorrelation with the specific fix for its worst failure. Plain autocorrelation matches a signal against a delayed copy of itself, and a delay of two periods lines up just as well as one — so it routinely reports a note an octave too low. YIN normalises by a running mean of the difference function, which suppresses those multiples, and then takes the first dip below a threshold rather than the deepest one.

Measured against pure tones from A2 to A5, this implementation lands within 0.7 cents — under one hundredth of a semitone. Against harmonically rich sawtooth waves, which are the classic octave trap, it makes no octave errors at all.

Turning pitches into notes

A pitch reading every ten milliseconds is not a melody. A note ends when the pitch moves by about a semitone, when the level drops, or when the detector loses confidence, and anything shorter than your minimum is discarded as a wobble at a note transition rather than something you played.

The pitch of each note is the median of its frames, not the mean. Attacks are messy — the first few milliseconds of a plucked string or a sung consonant can read wildly — and a mean lets those drag the note off. A median ignores them.

Velocity is scaled against the loudest note in your recording rather than an absolute level, so the dynamics of the performance survive. An absolute mapping puts every note of a normalised recording at 127 and flattens the phrasing completely.

Check and fix the notes before you download

The notes appear on a piano roll first, not straight in a file. A pitch tracker makes a few typical mistakes: a stray blip on a breath, one long note split in two by vibrato, a note that slipped an octave on a thin or breathy sound. Each takes a second to spot on a roll and far longer to hunt down in a DAW. Press Play notes to hear the transcription on a plain synth. Play with the recording plays both together, so a wrong note sticks out against the original.

Drag a note to move it in time or pitch, and drag its right end to change its length. Double-click empty space to add a note one beat long, and double-click a note to delete it. With a note selected, the arrow keys move it a semitone, or an octave with Shift, which is the quick fix for an octave slip. Snap to grid pulls every note onto the grid in one go, and Undo steps back through every change. The file you download is exactly what is on the roll.

Tempo and quantizing

Leave the tempo blank and the file is analysed with the same detector behind the BPM finder, and that tempo is written into the MIDI file so your DAW lays the notes on the right grid rather than assuming 120.

Snapping to a grid is off by default, and that is deliberate. Quantizing a hummed idea to sixteenths destroys the feel that made it worth recording, and it is trivial to quantize afterwards in any DAW but impossible to un-quantize. Turn it on when you are capturing something that was played to a click and you want it tidy.

Getting a clean result

• Record dry. Reverb and delay smear the pitch and create phantom notes at the tail of every phrase. If the recording has heavy effects, this will find notes that were never played.
• One take, one part. No harmonies, no double-tracking, no chords.
• Pick the range. Telling it you are working on a bassline stops it hearing a harmonic as the fundamental — the single most common source of a stray note an octave up.
• Trim the silence. Breaths and handling noise can register as very short notes; cutting the dead air first removes most of that.
• Humming works. A steady "ooh" or "ah" tracks better than "la la la", because consonants interrupt the pitch and produce extra note edges.

What the file contains

A standard MIDI file, type 0, one track, with your tempo and a 4/4 time signature written in. Note pitches, start times, lengths and velocities. It opens in Ableton, Logic, FL Studio, Reaper, GarageBand, MuseScore and anything else that reads MIDI.

It contains no sound. MIDI is a description of what to play, not a recording, so it will be silent until you point an instrument at it — which is the entire point: swap the sound, change the notes, transpose the key, fix the one wrong note by dragging it.