Move the note, not the clock

Why a naive pitch change sounds wrong

Take a single vowel and speed it up by a fifth. The pitch goes up, as intended, and the vowel also gets shorter - but a vowel is not one frequency. It is a stack of harmonics shaped by the throat and mouth into peaks called formants, and those peaks are what make an a sound like an a rather than an e. Resampling moves the harmonics and the formants by the same factor, so a voice that is raised an octave keeps its melody but loses the resonances that identify it. The result is readable as speech and obviously fake, which is exactly what people mean by the chipmunk effect.

There is a second, cheaper artifact: the splice. A time-domain algorithm cuts the signal into short frames and repositions them. If a frame is pasted where the waveform is at a different point of its cycle, the two halves cancel or add, and the seam becomes a click, a buzz, or a metallic flutter that gets worse the further you shift. This is why two pitch shifters with identical controls can sound nothing alike.

What this tool actually does

The signal is split into 2048-sample frames, half overlapping. For every new frame the algorithm searches a small window around the expected position for the offset that best matches the tail of the previous frame - a normalised correlation, computed on every fourth sample so the search stays fast enough for a phone. The frame is then added with a Hann window, and because the overlap is 50 percent the windows sum back to a flat line. That search is the whole trick: it chooses the splice point where the waveform was already going to be, so the join is inaudible instead of loud.

The stretched signal is then read back at the reciprocal ratio with linear interpolation between samples. Stretching by 1.335 and reading at 1.335 gives back the original length; the frequencies inside have been scaled by exactly 1.335, which is a perfect fourth. Do both channels with the same ratio and the stereo image survives, because the left and right channels are stretched with identical frame positions and the phase between them is preserved.

Choosing a number instead of guessing

For a voice that should sound like the same person, stay inside five semitones, and prefer the downward direction: lowering a voice adds weight and is hard to hear as processing, while raising it thins the tone and exposes the algorithm. Three semitones down is the everyday setting for narration, interviews and anything that will be listened to for an hour.

For music, work backwards from the key. If a sample is in A minor and the track is in C minor, the distance is three semitones - and playing it at +3 sounds more natural than nudging it by ear until it sits. If the sample has to move more than seven semitones, changing the arrangement is usually cheaper than the artifact you will spend an evening trying to hide.

For a joke, a character or a reveal, an octave is the strongest interval available and the one that reads instantly as a deliberate effect. Use it whole; three or seven semitones in a comedic context sound like a mistake, while twelve sounds like a decision.

What to listen for before you export

Listen to the quiet parts first. A seam that is invisible under a loud chorus often shows up in a breath, in the decay of a piano note, or in room tone between sentences - those are the places where the algorithm has no periodicity to lock onto and where a wrong splice is most exposed. If the tail of a word flutters, lower the shift by a semitone or two and compare.

Then check the duration against the source. It should be identical to within a sample or two; a difference of a tenth of a second means something in the chain resampled the file, which will drift against a video. Finally, compare loudness rather than peak: shifting changes the phase relationships inside the signal, so a file that peaks at the same level can measure a decibel quieter.

Formats and generations

Choose WAV when the result will be edited further or uploaded to a video editor, and MP3 when it is going straight to a device or a chat. Shifting an MP3 that was already shifted is a second generation of lossy encoding on top of an algorithmic artifact - if you need to try several settings, keep one WAV and re-run the shift each time rather than shifting the shifted file.

Nothing here is uploaded. The file is decoded by the browser's own audio engine, processed in the tab and handed back as a download, which is why the tool works offline once the page has loaded - and why a fifteen-minute file is the practical limit, set by memory rather than by policy.