Audio Pitch Shifter

Keeping the length while moving the pitch is done with WSOLA (waveform similarity overlap-add) followed by a resample of the reciprocal amount. The stretch slides each output frame to the position where it best continues the previous one, which is what keeps periodic material such as a sung note or a bowed string from developing a stutter, and the resample then puts the duration back where it started. Both halves come from the same engine the Voice Changer uses. Across ten intervals from a full octave down to a full octave up, the worst error is 0.58 cents with the length held to within a thousandth of a percent; driven through the browser and measured back out of the exported file, a perfect fifth reads -0.22 cents and an octave +0.47 cents.

This replaced a grain loop that advanced its read and write pointers by the same hop, so the four overlapping grains covering any output sample read the source at four different phases and largely cancelled. The damage was not subtle: asking for +7 semitones produced 698.4 Hz from a 440 Hz source instead of 659.3 Hz, which is 99.8 cents sharp, nearly a whole semitone wrong, 15 dB quieter than the input, and with more noise than signal. Asking for -5 landed 122.8 cents sharp. The vari-speed path had a separate fault: it relabelled the buffer sample rate rather than resampling the audio, then handed the encoder a rate no MP3 file can carry, so the exported file came back declaring 48 kHz for a 44.1 kHz source.

Vari-speed mode is now a genuine cubic resample that leaves the sample rate alone, so pitching up shortens the file and pitching down lengthens it, exactly as a tape machine would, and the file still declares the rate it actually has. The preview plays the processed audio in both modes, so what you hear before exporting is what the file contains; the old preview always used the browser playbackRate, which meant the length-preserving mode sounded like vari-speed right up until you downloaded something different. The stretch runs in a Web Worker with a progress bar and a Cancel button, since it is a full pass over every sample: a three minute stereo file takes a few seconds and the tab stays responsive throughout.

The Measured Result panel is the check. After each render the output is analysed by a short-time Fourier transform that knows nothing about how the audio was made, and the panel reports the source pitch, the target, what was actually measured, and the error in cents. On clearly pitched material that number should sit within a cent or two of zero. On dense or percussive material the strongest partial is not always the note, and the panel says so rather than pretending otherwise. There is still no formant correction, so large shifts move vocal-tract resonances with the pitch and take on the familiar chipmunk or deep character.