Audio to MIDI

Hum one comfortable note and hold it for two beats. Now hum the same note as four short notes in the same two beats. Same pitch, two completely different phrases: one is a held sigh, the other a little stutter.

Here's a sung line to listen to. Follow it in the spectrogram: each bright horizontal band is a pitch, and each break or bend in a band is a decision someone has to make about where one note ends and the next begins.

Before reading on: if a pitch detector listened to your two hums, what would it report? It might name the same pitch, correctly, both times. That label alone wouldn't tell you whether you sang one long note or four short ones. Turning audio into notes is two jobs, and the harder one is deciding where notes start and stop.

Two decisions in every note

A recording is a continuous signal; a MIDI part is a list of separate events. To get from one to the other, a transcription tool has to decide, moment by moment:

• Is there a pitched note here, and what is it? That's pitch estimation.
• Where does this note start and end, and is this a new note or the same one carrying on? That's segmentation, finding the onsets and offsets.

Pitch gets most of the attention, but segmentation is what makes a phrase sound right. A breath before a note, a slide up to it, a little vibrato, a repeated note with no gap: each of those asks "one note or two?" and there's often no single correct answer.

Fix the draft

Here's that draft, played as notes. Press play and you'll hear the stutter where the long G should be, and the last note leap up. Then use the control to move note 6 down twelve semitones.

The octave fix took one click and every pitch is now right. The phrase is still wrong: the long, singing G you meant is a nervous triple. A transcription can have every pitch correct and still be the wrong phrase.

Why tools make these mistakes

Most errors come from things that are musical to you and ambiguous to a machine:

• Octave errors. A low note often has strong upper harmonics, sometimes stronger than the fundamental, so the detector picks the octave above.
• Split notes. Vibrato, a wobble in level, or a breath in the middle of a held note can look like a new attack.
• Merged notes. Two repeated notes at the same pitch with no gap between them can look like one long note.
• Slides become stairs. A voice or guitar slide glides through pitch continuously. A note list can only store steps, so a slide becomes a run of short notes or one note at the wrong moment. Pitch bend could hold the curve, but most note-only transcriptions throw it away.
• Tails become notes. Reverb or a ringing string can keep a pitch alive after you stopped, so the note ends late.

Modern tools are good. Open-source transcription models such as Spotify's Basic Pitch (2022), and note-editing pitch software, have made this everyday studio work. But they are estimating what you meant from what you recorded, and the estimate is a draft.

Check rhythm before pitch

When a draft comes back, it's tempting to scan for wrong note names. Do it the other way round:

1. Count the attacks. Listen to the original once and tap every new note. If you sang four and the draft has nine, look for splits. If it has three, look for merges.
2. Check where notes start. A right note arriving a little late can drag the whole groove.
3. Check where notes end. A note that runs into the next one turns a clean phrase into mush on a piano sound.
4. Then check pitch, starting with big jumps like octave errors.

This keeps your attention on the phrase instead of a forest of small labels.

You meant	The draft shows	Try	Then you should hear
One held G	Three short Gs	Merge them into one note	One long, continuous tone
Two separate Cs	One long C	Split it where the second attack is	Two distinct attacks
A quiet first D	Nothing	Add the note	The phrase starts where you sang it
E4	E5	Move it down an octave	The line falls back where it belongs

Keep the feel you wanted

Once the obvious errors are fixed, you'll be tempted to quantize: snap every note to the grid. Sometimes that's right. A sketch that will sit on a programmed beat often wants to be tight.

Snapping needs a grid. If you recorded without a click, the tool has to find the beat first, and Beat and tempo tracking explains why it may hand you double or half the tempo you felt.

But your hum had a feel. Maybe the last note landed lazily late, or the first one anticipated the beat. Those are the things that made it yours. Quantize a copy, compare, and keep whichever serves the song. Correct a timing mistake; don't correct a timing choice.

Velocity deserves the same care. The tool's guess at how loud each note was isn't automatically the right touch for a piano or a synth. Set accents once you know which instrument will play the part.

Same idea, different place: pitch correction has to make the same "which note did you mean?" call, but live, without hearing what comes next. See Pitch shifting.

When the source is the problem

Some recordings are hard for any tool: a mumbled hum, a line sung over a backing track, a guitar with lots of reverb. Before you fight the transcription, go back and listen. Is there a clear attack? Is there really one pitch at that moment, or two overlapping?

For a short idea, typing six notes in by hand can be faster than repairing a messy draft. Use transcription when it saves time or catches something you'd have missed, not because it's there.

Make it yours

Record a short melody of your own on your phone, hummed or played on one instrument, one note at a time, with a moment of silence before it. Run it through the audio-to-MIDI feature in your DAW or an open-source transcription tool.

Then fix the draft in the order above: attacks, starts, ends, pitch. Play your corrected notes on FM Keys and compare with the recording. Save three things together: the recording, the raw draft and your corrected version. Write one sentence about a decision you made ("I kept the late last note"). That sentence is your taste, written down, and it's what you'll want to remember when you build the idea into a song.

What comes after the corrected melody? Hum to arrangement carries it into chords, a bass part and two contrasting sections, while keeping the phrase that made the original hum worth saving.

Quick check

1. A transcription has every pitch right, but your melody sounds nervous and choppy when played back. What's the most likely cause? Segmentation: held notes split into several short ones, or notes ending too early. Pitch was only half the job.
2. A low bass line comes back with several notes an octave too high. Why do tools do this? Low notes often have strong upper harmonics, so the detector locks onto the octave above. It's one of the most common errors and one of the easiest to fix.
3. You sang a slow slide from G up to C. The draft shows G, A, B, C as four short notes. Is the tool wrong? It did what a note list can do. A slide is continuous and notes are steps. Decide whether you want one note, a few steps, or a pitch bend, based on the instrument that will play it.

Next steps

• Go deeper: Composing with a sampled library. Your tune is notes now. Which instrument should play it, and how do you write a part that suits that instrument?
• Use it: What an audio model learns. Why a model's useful output still needs a musical ear.
• Sideways: Chord and key recognition. The same "what did you mean?" problem for harmony, where one sound can honestly carry two chord names.