@audio/denoise-dewow

Wow & flutter correction — pitch-drift removal for tape/vinyl/cassette transfers

npm install @audio/denoise-dewow

import dewow, { analyze } from '@audio/denoise-dewow'

Estimates the transport's instantaneous speed over time and corrects it by variable-rate resampling — the classical (non-ML) counterpart of Celemony Capstan. Three estimators: track the phase-vocoder instantaneous frequency of stable spectral partials ('partial', default — McAulay & Quatieri 1986 partial tracking + Godsill & Rayner, Digital Audio Restoration, 1998, ch. 6), lock onto one known tone such as mains hum or a calibration tone ('reference' — Czyżewski et al., Wow detection and compensation employing spectral processing of audio, JAES 2007), or track monophonic pitch ('pitch' — @audio/pitch-pyin). See Howarth & Wolfe, Correction of Wow and Flutter Effects in Analogue Tape Transfers, AES 117th/118th Convention, 2004/2005, and Nichols, The Digital Restoration of Wow and Flutter Distorted Gramophone Recordings, 1999.

let corrected = dewow(recording, { fs: 44100 })                       // default: track partials
let corrected = dewow(recording, { fs, mode: 'reference', refFreq: 50 }) // lock onto 50 Hz mains hum
let corrected = dewow(recording, { fs, mode: 'pitch', smooth: 3 })    // monophonic voice/instrument

let meter = analyze(recording, { fs })
// → { speed, times, hop, wow, flutter, wowPeak, flutterPeak, confidence, tracks? }

recording is a mono Float32Array or an array of channels ([L, R, …]); analysis always runs on the mono mix, and one shared curve corrects every channel, so a stereo pair stays sample-aligned. Returns new arrays — never in place.

| Param | Default | | |---|---|---| | fs | 44100 | Sample rate | | mode | 'partial' | 'partial' | 'reference' | 'pitch' | | refFreq | — | Known tone/hum frequency (Hz) — required for mode: 'reference' | | frameSize | 4096 | STFT frame for peak-picking / track continuity | | hopSize | 512 | STFT hop — also the per-hop curve's own sample rate (its Nyquist, fs/(2·hopSize), caps flutter frequency) | | smooth | 0.05 | Zero-phase smoothing time constant (s) separating wow (below) from flutter (residual) | | wow | true | Correct the low-passed (<~6 Hz) component | | flutter | true | Correct the residual (faster) component | | maxDeviation | 0.05 | Clamp the corrected speed ratio to [1−x, 1+x] | | minTrack | 0.5 | Shortest partial kept, in seconds — 'partial' mode only | | minFreq / maxFreq | 50 / 2000 | f0 search range — 'pitch' mode only | | keepLength | true | Output length equals input length |

analyze() is the estimator alone — the "wow & flutter meter" — with no audio output. wow/flutter are the unweighted RMS deviation in %; wowPeak/flutterPeak are the unweighted peak deviation in %. These are not the IEC 60386 / DIN 45507 figure, which applies a psychoacoustic weighting curve (peaking near 4 Hz) before measuring — that weighting filter isn't implemented here, so a reading from this meter isn't directly comparable to a spec-sheet wow-and-flutter number. confidence is the fraction of hops with a usable estimate (a stable track, a present reference tone, or voiced pitch); tracks (mode 'partial' only) lists the accepted partials.

Resolution trade-offs, measured, not assumed:

• Flutter ceiling. The per-hop curve is itself sampled at fs/hopSize (86 Hz at the defaults) — anything above its Nyquist, fs/(2·hopSize) ≈ 43 Hz, can't be represented at all. Below that, the phase-vocoder estimate needs the analysed signal to be roughly stationary across its whole analysis window, not just one hop — a 4096-sample window (93 ms) averages away a 30 Hz flutter almost entirely (measured: a synthetic ±0.4% 30 Hz flutter reads back as ~0, correlation −0.18 against the true curve). 'partial' mode works around this with a second, short window (1024 samples, 23.2 ms) purely for the frequency reading, while the long window still does peak-picking / track continuity (which needs the frequency resolution to keep this repo's own test chord's 220/330/440/660 Hz partials from crosstalking). That recovers 30 Hz flutter to within the tolerances below, but a shorter hopSize is the honest fix for tracking flutter closer to the nominal 100 Hz ceiling.
• 'reference' mode and low reference frequencies. A 50/60 Hz mains hum needs a much longer reading window than a musical partial does — one cycle of 50 Hz is 20 ms, comparable to the flutter-tracking window above, so 'reference' mode scales its window to the target frequency (~4.5 cycles) instead of using the fixed one. It also band-passes (Q 5) around refFreq before reading phase: program content sharing that band (speech has real energy down at 50 Hz) otherwise corrupts the reading (measured: 50 Hz hum at −30 dB under 6 s of speech, curve correlation 0.94 unfiltered → 0.97 filtered).
• 'pitch' mode and real speech/instruments. smooth doubles as the mode's own vibrato/prosody-vs-drift cutoff (speed = f0 / lowpass(f0, smooth)): the default 0.05 s tracks pitch fast enough that genuine slow wow gets absorbed into the baseline and cancels out of the ratio, leaving only fast flutter as a visible deviation. A smooth of several seconds recovers slow wow, at the cost of also absorbing real vibrato/prosody as if it were flutter — natural speech intonation moves far more (measured: tens of percent) than a 2% wow defect, so this mode is a functional correction, not a precision one, on unpitched or lightly-inflected material. It's also frame-independent (@audio/pitch-pyin has no Viterbi/HMM smoothing across frames — see that package's own README) — no octave-jump correction is applied.

Correction. Once the speed curve is built, correction is a single windowed-sinc read (@audio/resample-sinc, 16 zero-crossings) per output sample at a warped position — pos[n] = pos[n-1] + 1/speed(pos[n-1]), narrowing the anti-alias cutoff whenever the local read rate exceeds 1×. Measured on the synthetic 2%[email protected] Hz + 0.4%-flutter-@30 Hz defect used in the tests: the 440 Hz partial's own residual frequency deviation drops from 1.4% to 0.18% RMS (target ≤0.2%); a clean, undistorted input passes through at ≥40 dB SNR (the estimator finds speed≈1 and a sinc read at near-integer positions is near-identity). A plain sample-domain SNR against the pristine original is not meaningful after correction, even a mathematically exact one — see the source comment on localSnr in test.js for why, and use analyze()'s own residual-deviation-style check instead if you need a number.

Not implemented: azimuth/head-alignment error (a frequency-independent time skew across the stereo image, not a speed error — out of scope), dropout/gap repair (@audio/denoise-repair), or anything ML-based. Celemony Capstan and similar tools additionally use trained models to separate genuine musical vibrato from mechanical wow on program material with no stable partial or reference tone at all; this package only ever measures speed from spectral evidence actually present in the signal.

Use when: tape hiss/wobble on cassette or reel-to-reel transfers, turntable speed instability (belt/motor wear), 16mm/optical-track flutter — anything with a stable tone, a mains-hum residual, or several seconds of sustained pitched content to lock onto. Not for: dropouts, azimuth error, or program material with no stable partial and no reference tone (a cappella breath, pure noise, hard cuts) — the estimator has nothing to track.

Part of @audio/denoise — the denoise family umbrella.

MIT © audiojs