Change the key. Without changing the tempo.

At a glance

What actually changes

Pitch and speed are naturally coupled. Play a record faster and it gets higher; slow it down and it gets lower. Separating them takes two steps that cancel each other's side effects.

First the audio is resampled, which shifts pitch and tempo together — exactly the record-player effect. Then it is time-stretched by the inverse amount, which puts the tempo back where it started while leaving the new pitch alone. The net result is a file the same length as the original, in a different key.

The time-stretch is the hard part, and it is done with WSOLA — an overlap-add method that finds the point of best waveform correlation before splicing, so the joins land on matching phase rather than creating clicks. It is the same well-tested implementation used by the speed tool, which does the same job in the opposite direction: tempo without pitch.

Shifts are measured in semitones because that is how music works. Each semitone is a ratio of the twelfth root of two, so twelve of them make exactly an octave — a doubling or halving of frequency.

When you'd want this

• Transposing a backing track into your range. The most common use by a distance — a song sits a couple of semitones too high to sing comfortably, and moving it down two or three fixes it without making the track drag.
• Matching two songs for a mashup or a DJ transition, where the tempos already line up but the keys clash.
• Fixing a recording made slightly out of tune — a tape transfer running fast, or an instrument that drifted. One semitone, or a fraction of one, brings it back to concert pitch.
• Practising an instrument by moving a piece into an easier key without slowing it down.
• Voice disguise or character work, where a few semitones down changes the character of a voice while keeping the delivery intact.

Settings that matter here

Stay within about ±5 semitones for the best results. Small shifts are close to transparent. Beyond a fifth the artefacts of time-stretching start to become audible — a slight smearing on transients, a hollow or metallic quality on sustained sounds — because the algorithm is having to invent progressively more.

Octave shifts (±12) work, and are useful for effects, but do not expect them to sound natural. Anything that far from the original is a deliberate transformation rather than a transposition.

Voices tolerate downward shifts better than upward ones. Moving a vocal up tends to produce the chipmunk quality even with tempo preserved, because the formants — the resonances that make a voice sound like a particular size of person — move with the pitch. Downward shifts sound more natural.

Output at 320 kbps or WAV if the result is going anywhere further. Pitch shifting is a re-encode on top of whatever the source already was, so give the encoder room.

What you lose

A little fidelity, unavoidably. Time-stretching reconstructs the signal rather than copying it, so transients — drum hits, plucked strings, consonants — soften slightly. At small shifts this is very hard to hear; at large ones it is obvious.

Formant realism on voices. Because the whole spectrum moves together, a shifted voice sounds like a different sized person rather than the same person singing higher or lower. Formant-corrected shifting is a more complex process than this does.

Processing time. The file goes through the WebAssembly build of ffmpeg, which downloads on first use and is cached afterwards. A long file takes real time to process.

If you want the opposite operation — same pitch, different speed — that is the speed tool, not this one.