Browser-side vocal pitch correction for home audio recordings. Drop in an audio file (voice over a backing track, or a solo vocal), pick a key — or let it detect one — and download a pitch-corrected WAV. Runs entirely in the browser. No server, no uploads, no account.
https://bubiche.github.io/pitch_correction/
Tested on desktop Chrome and Firefox. Mobile and Safari are best-effort
— OfflineAudioContext quirks and memory limits make them unreliable
for full-mode separation.
Two modes:
- Direct — runs YIN pitch detection plus a WSOLA pitch shifter (soundtouchjs) on the whole signal. Best for solo vocal recordings. Will produce artifacts on polyphonic input.
- Full — separates the vocal from the accompaniment with an MDX-Net ONNX model running in the browser, pitch-corrects the vocal stem only, then mixes the corrected vocal back over the untouched accompaniment. Best for full songs with one singer over a backing track.
Both modes snap detected pitch to the nearest scale tone in the selected key. The retune speed slider blends between passthrough and hard-snap (the classic Auto-Tune sound).
Other niceties:
- Auto key detection runs on drop (Krumhansl–Kessler) and pre-fills the key selector; override it if you disagree.
- A/B playback — listen to the original and the corrected output back-to-back without re-downloading.
- Pitch contour visualization — overlays detected pitch, snap target, and corrected pitch over time.
- WebGPU when available, threaded WASM otherwise. A 2:30 song separates in ~60–90 s on an M4 Max GPU, ~5 minutes on threaded WASM.
Pinned to pnpm 11 and Node 24 via .tool-versions.
pnpm install
pnpm dev # local dev server
pnpm typecheck # tsc -b
pnpm test # tsx --test on src/lib/**/*.test.ts
pnpm build # production bundle to dist/
pnpm preview # serve dist/ at the /pitch_correction/ base
pnpm fixtures # regenerate diagnostic WAV fixtures for the pitch shifterThe separation model (~64 MB) is committed under public/models/ so it
ships with the Pages deploy. First load fetches it once and caches it
in IndexedDB; subsequent runs read from cache.
See ARCHITECTURE.md for the threading model,
audio pipeline, library choices, and the click-suppression chain on the
pitch shifter.
See CREDITS.md for library, model, and dataset
attributions.
AGPL-3.0. The project bundles pitchfinder (GPL-3.0) for
pitch detection, which forces strong copyleft on derivatives anyway —
AGPL extends that to the hosted-app case so users of the public site
have the same rights as users of the source.