How does Auto-Tune fix a singer's voice?

Auto-Tune fixes a voice in two steps: it works out what note the singer is actually producing, then nudges it to the exact pitch of the nearest proper note. To find the note, it uses a trick called autocorrelation, comparing the sound wave with a slightly delayed copy of itself to find how often the wave repeats. That repetition rate is the pitch. If the singer is a little flat, the software shifts the sound up to the right note.

Used gently, it is practically invisible. Its inventor, Andy Hildebrand, says a lightly corrected note is impossible even for skilled producers to detect. Turn the correction speed to the extreme, though, and the voice stops sliding naturally between notes and jumps from note to note like a synthesizer. That is the famous robotic sound, first used on purpose on Cher's 1998 hit "Believe".

The backstory is wonderfully unlikely. Hildebrand was a trained flutist who spent years in the oil industry, processing seismic data to find oil underground, using autocorrelation on huge datasets. One day, a colleague's wife joked that he should build a box to make her sing in tune. He realized his oil-hunting math could do exactly that, and Auto-Tune was released in 1997.