How Does Autotune Work? Pitch Correction Explained for Producers

Most producers know what autotune sounds like long before they know what it's doing. You drop it on a vocal, pick a key, turn a knob, and either it sounds like a better take or it sounds like a robot. When it goes the wrong way, it's hard to fix a process you can't picture.

So this is the picture. Everything below uses our own plugin, TUNE, because we can show you exactly what it does inside, but the three steps are the same in every pitch corrector you'll come across. Once you know them, the knobs on all of them start to make sense.

The examples are three seconds of White Flag by CODA, a real vocal from the catalog, in F# major. The traces aren't drawings. They were measured from the audio that came out of the plugin, 100 points a second.

The three jobs

Strip away the interface and a pitch corrector runs a loop:

1. Measure. What pitch is the voice on right now?
2. Decide. Which note in the key was the singer going for?
3. Move. Shift the audio by the difference, at the speed you've set.

TUNE runs that loop roughly every 6 milliseconds, about 170 times a second. Each step has its own problems, and nearly every control on a tuner exists to deal with one of them.

Look at how little of that line sits exactly on a note. That's normal. Good singers scoop into notes, drift a few cents, and hold long notes with vibrato. The question a tuner has to answer is which of those movements are mistakes and which are the performance.

Step 1: measuring the pitch

A sung note is a repeating waveform. The pitch you hear is how many times per second that shape repeats: A above middle C repeats 440 times a second, so it's 440 Hz. Each semitone up is about 6% faster, and each semitone is divided into 100 cents, which is the unit you'll see in every tuner.

To find the pitch, TUNE compares the incoming audio with a delayed copy of itself. When the delay is exactly one cycle long, the two copies line up almost perfectly, and that delay is the period of the note. The method is called YIN. It's been around for over twenty years and is still one of the most common ways to track a single voice, because it's accurate and doesn't need much audio to work with.

Octave errors

There's a catch. A waveform that lines up with itself after one cycle also lines up after two. Occasionally the best match is the double-length one, and the tracker reports a note an octave too low. You've heard this if a tuner has ever briefly thrown a note down an octave on a breathy passage.

TUNE handles it by scoring several candidate periods rather than trusting the first match. Anything that still jumps an octave away from its neighbours is pulled back toward the rest of the note instead of being passed straight to the shifter.

Why low voices need more time

To recognise a cycle, the tracker has to see at least a couple of them. A low male voice around 80 Hz has cycles 12.5 ms long, a soprano's are a quarter of that. The lowest note the tracker has to be ready for sets how much audio it needs before it can answer. TUNE Pro has a voice type setting for this: telling it roughly where the singer sits narrows the range it has to search.

S's, T's and breaths

Consonants and breaths don't repeat. An S is noise. It has no pitch, so there's nothing to correct, and a tuner that tries anyway produces the lispy, zippery artefacts people blame on autotune.

TUNE decides, frame by frame, whether the voice has a pitch at all. Where it doesn't, the audio crossfades back to the untouched signal. S's, T's and breaths come through exactly as they were recorded. In TUNE Pro, Gate sets how strict that decision is: turn it up and less of the take gets pitched, more passes dry. Floor is a separate noise gate that fades the output to silence below a level you set, which helps when there's room noise or headphone bleed on the track.

Step 2: deciding the note

This is where tuners differ most, and it's the step that decides whether a vocal sounds tuned or sounds processed.

The key comes first

Setting the key and scale gives the tuner its list of allowed notes. In F# major that's F#, G#, A#, B, C#, D# and F (strictly E#, but every DAW calls it F). TUNE shows the list right under the key selector, so you can see what it will tune to.

TUNE doesn't detect the key for you, and that's deliberate. We built key detection, tested it on vocals with a known key, and it got half of them wrong. It isn't fixable from a vocal alone: a melody that never sings a particular scale degree fits two keys equally well, and the key finders that work listen to the whole mix, which a plugin on the vocal track never hears. Setting it yourself takes five seconds, and guessing it wrong makes every note wrong.

Why snapping to the nearest note doesn't work

The simplest tuner takes each pitch measurement and snaps it to the closest allowed note. It works on a steady, well-sung note. It falls apart on vibrato.

Vibrato is the pitch swinging above and below a note five or six times a second, often by 40 or 50 cents either way. Snap every measurement to the nearest note and one of two things happens. Either the whole wobble gets flattened into a dead, straight line, or, if the wobble crosses the halfway point between two notes, the tuner flips back and forth between them and you get a warble that was never sung.

Note centre plus vibrato

So TUNE doesn't correct the raw pitch. It splits it into two parts: the centre of the note the singer is holding, and the vibrato moving around that centre. Only the centre is moved onto the scale. Then the vibrato is put back on top, by as much as the Vibrato setting says.

Finding the centre properly is the hard part. TUNE starts measuring at the moment a new note begins, ignores anything more than a semitone away from where the voice currently is, and when the pitch is clearly oscillating it measures how long one vibrato cycle takes and averages over whole cycles. That puts the centre in the middle of the wobble rather than wherever the wobble happened to be when the note started. In testing, Natural kept 98% of a vibrato 50 cents deep, while Hard Tune kept 14%.

This split is what the presets are really setting. Natural keeps all of the vibrato, Modern Pop keeps about half, Hard Tune keeps none. The correction of the note underneath is the same in all three.

Notes that sit right between two others

A voice hovering near the midpoint between two notes is the worst case for any tuner, because a few cents either way changes the decision. That's where you hear notes flip. Most of the time the note centre solves it, but SAUCE, which snaps every movement, needs something extra. It only switches note when the voice is clearly past the halfway point and the audio just ahead agrees, and a note held inside that grey zone is allowed to settle for about 120 ms before it moves. The result is the full hard sound without the random jumps.

What Pro adds to the decision

The Advanced controls in TUNE Pro all adjust this step:

• Humanize gives held notes their natural movement back while short notes still snap. On notes held longer than about 120 ms, the amount of vibrato kept rises until it's all back by around 400 ms. That's how a hard-tuned hook can still sound sung.
• Flex-Tune makes the correction weaker the further the voice is from a note. Slides and blue notes, sung well away from the scale, are left alone. Pitch sung close to a note is still corrected.
• Strength sets how far toward the target the correction goes.
• The note keyboard removes notes from the scale. If a slide between two notes keeps catching a third on the way, take that note out and TUNE goes straight past it.
• Wobble is the hard-tuned trill: the target alternates between the corrected note and the next note up in the scale, about 5.4 times a second.

Step 3: moving the voice

Retune speed

Once the tuner knows how far off the voice is, it has to decide how quickly to fix it. That's retune speed.

At 0 ms the correction arrives instantly. Every note snaps the moment it starts and every slide between notes becomes a jump, and that's the stepped sound people mean when they talk about hard autotune. At 42 ms, where the Natural preset sits, the correction eases in over the start of each note, the way a singer lands on a pitch. TUNE's range runs up to 200 ms, which only touches notes that stay off for a while.

One detail matters a lot here. In TUNE, retune only slows down the correction of the note. It never slows down the vibrato, which is put back separately. If both went through the same smoothing, a slow retune would blur the vibrato into a lazy wobble, and a fast one would chase it. Keeping them apart is why Natural can be quick on the note and still leave the vibrato alone.