@audio/stretch-wsola

Waveform-similarity overlap-add (Verhelst & Roelands 1993), in the segment form of SoundTouch and sox tempo. Segments of the input are laid one hop apart, each crossfaded into the one before, each read where its head best continues the one before (normalized cross-correlation within ±delta of where the time map puts it). The waveform is copied, never resynthesized: a voice keeps its glottal pulse shape and its consonants their attacks, which a phase vocoder smears into a distant, reverberant sound.

import wsola from '@audio/stretch-wsola'

let out = wsola(samples, { factor: 1.5 })
let [l, r] = wsola([left, right], { factor: 1.5, sampleRate: 48000 })   // one search for both: the image stays

let write = wsola({ factor: 2 })   // streaming
write(block1)
let tail = write()                  // flush

| Param | Default | | |---|---|---| | factor | 1 | Time stretch ratio, or a function of input seconds | | sampleRate | 44100 | | | frameSize | 40 ms | Segment | | hopSize | frameSize · 3/5 | Output hop: segments cross over the rest, 16 ms (a 62.5 Hz period) | | delta | 8 ms | Search, ± (half that period) |

Hosts with their own time map use the engine: stretcher(channels, { at: s => inputSample, end, sampleRate }) → { write(channels), end() }, its output aligned to its sample 0; where the map runs at rate 1 the output is the input. end, the input's length once known (a number or a function), keeps the last segments reading inside it.

The crossfade follows how alike the two sides are (Fink, Holters & Zölzer 2016): linear for the same waveform, equal-power for noise.

On speech (VoiceBank, Spoken Wikipedia, a lecture), ×0.7 to ×2: the spectrum 0.8 to 1.8 dB closer to the input's than the previous version (Hann frames at ¾ overlap), within 0.25 dB of sox tempo -s; 0.91 to 1.08 of the input's pulse peakiness kept, a phase-locked vocoder 0.62 to 0.71.

Use when: shortening speech, a solo voice, a monophonic instrument. Slowing, its repeats turn noise and reverberation into a flanger: use @audio/stretch-pvsola. Not for: chords and mixes: one alignment cannot fit several pitches (a chord's balance at 2×: 0.18, the phase-locked vocoder's 0.9). Use @audio/stretch-pvoc-lock or @audio/stretch-transient.

Part of @audio/stretch.

License

MIT · ॐ