Pitch shifting and correction

Press play. A singer, shifted up an octave. The melody is right, the timing is right, and yet it isn't a higher version of the same person. It's a cartoon. In 1958 Ross Bagdasarian built a whole act on this sound: he sang slowly into a tape machine, sped the tape up, and The Chipmunks were born.

Something besides the pitch changed when the voice went up. What is it? Is it something about the notes, or something about the voice?

Two things live in a voice

When you sing a note, your vocal folds buzz at a steady rate. That rate is the pitch: 220 times a second for the A below middle C. The buzz is full of harmonics, at 440, 660, 880 Hz and up, all whole-number multiples of the note.

Then your throat and mouth shape that buzz. They act like a set of resonant chambers that make some frequency regions louder and others quieter. Those boosted regions are called formants, and they're what make an "ah" sound different from an "ee". Try it: sing "ah" and slide into "ee" on one note. The pitch stays put. The formants move.

Formants also tell your ear how big the singer is. A small head and a short throat put formants high; a big one puts them low. That's the clue the lab just broke.

Why the simple shift makes a chipmunk

The shifter in the lab works much like speeding up tape, except that it keeps the timing. It chops the voice into tiny overlapping slices, about 50 ms each by default, and plays each slice back faster, crossfading so the seams are hidden. Every frequency gets multiplied by the same amount, including the formants.

So when the note goes up an octave, the vowel shape goes up an octave too, and your ear hears a head half the size. Small shifts get away with it. Two or three semitones sounds like a brighter, younger version of the singer; that's why gentle shifts make convincing harmony doubles. Large shifts don't.

The Window control sets how long those slices are. Short windows keep the timing tight but make the voice flutter and buzz. Long ones smooth it out but blur consonants and start to sound echoey. Every pitch shifter makes a trade like this somewhere; this one lets you hear it.

Move the note and the vowel separately

To avoid the chipmunk, a shifter has to treat the two ingredients separately: move the harmonics to the new note, but leave the formants where they were. That's formant preservation. The lab below makes the split easy to hear because it rebuilds the voice from a synth: it follows the singer's notes and vowel shapes, then plays them back with a sawtooth wave.

Raise Octave and the melody jumps up while the vowels stay put: higher, but no smaller. Put it back and raise Formant Shift instead, and the notes stay where they were while the voice shrinks. Lower it and the same melody comes from a giant. Two knobs, two different jobs. A formant-preserving shifter does the first move on the real voice, with no robot in between.

It's never perfect. When a real singer goes up an octave, they also change how they breathe, how they shape vowels, how hard they push. A preserved-formant octave still sounds a bit like a singer who never practised up there. Within a few semitones it can be very convincing. Every shift here still moves your own voice. Making it sound like a different person takes a model trained on that person's recordings, and Voices, cloning and consent covers what such a model keeps and whose permission it needs.

Correction: choosing where each note lands

Pitch correction, first sold as Antares Auto-Tune in 1997 and a sound of its own after Cher's "Believe" in 1998, adds a decision. A shifter moves everything by a fixed interval. A corrector listens to each moment, works out which note the singer is aiming for, and nudges the voice onto it.

The lab above does a version of this too. It snaps every note it hears to the nearest semitone before playing it. Retune Speed decides how fast. Slow, and the scoops and vibrato of the singing survive; the voice glides onto each note. At zero, every note jumps instantly into place, and the movement between notes turns into steps. That's the sound Cher's "Believe" made famous in 1998, and that a whole generation of R&B and hip-hop turned into an instrument. Try Retune Speed at 0, then at 150 ms, and listen to the slides between notes.

The most important choice in correction isn't the speed, though. It's the target. A corrector set to the wrong key will push a singer confidently onto the wrong note, and a faster retune just gets them there sooner. Choose the scale before you choose the speed.

Where the detector gets confused

Correction and the robot lab both depend on a pitch detector, and detectors expect one clear note at a time. They stumble in predictable places:

• Breath and consonants. An "s" or a whisper has no pitch to find.
• Chords and doubles. Two notes at once give the detector two answers, and it may jump between them.
• Octave slips. A strong harmonic can fool it into hearing the note an octave too high or low.

That's why correction works best on a dry, solo vocal, before reverb and harmonies are added.

Same idea, different place: turning a hummed tune into MIDI notes uses the same kind of detector, and it slips an octave in the same places. See Audio to MIDI.

Make it yours

Take a vocal of your own and decide on one job. A double: a copy shifted by a third or a fifth, mixed quietly under the lead, using a small shift so it still sounds human. Or a correction: set the key first, then find the slowest retune speed that still lands the notes you want. Or go the other way and use a hard snap as a deliberate effect.

Save the setting with its job and key in the name ("fifth-up double, A minor") so the next time you write in that key, the harmony is one click away.

Quick check

1. The shifted vocal is in tune but sounds like a child. What moved that shouldn't have? The formants. The shift multiplied the vowel resonances along with the notes. A formant-preserving shift, or a smaller interval, keeps the singer's size.
2. You shorten the retune speed and the singer lands on a wrong note faster. What's the real problem? The target. The corrector is doing exactly what it was told; it's aiming at the wrong scale. Fix the key or the allowed notes first.
3. You try correcting a stack of backing harmonies in one go and it warbles. Why? The detector expects one note at a time. With several notes it can't settle on one, so it hops between them. Correct each voice on its own.

Next steps

• Go deeper: Frequency shifting and ring modulation. A pitch shifter multiplies every frequency by the same amount. What happens if you add the same number of hertz instead?
• Use it: Chords and voicing. Choose the intervals for a harmony double so they fit the chords underneath.