@audio/measure-correct

Response correction design — measured response + target curve → FIR and/or parametric correction EQ

import correct from '@audio/measure-correct'

Room/headphone/speaker linearization, REW/Dirac/Acourate class: given a measured response (an impulse response, a {freqs, db} curve, or {f, gain} points) and a target curve, designs a correction EQ that flattens the difference. The measured curve is fractional-octave smoothed, level-aligned to the target, then inverted with Kirkeby-regularized deviation (correction = dev·|H|²/(|H|²+λ) in linear magnitude — a deep null in the measured response gets rolled off toward 0 rather than boosted to cancel it, since a literal 1/H inverse blows up exactly where H is small) and clamped to boost/cut limits, tapering to 0 dB outside [fMin, fMax]. The correction renders as a linear-phase (or minimum-phase) FIR via @audio/eq-fir's frequency-sampling design and/or as parametric bands via @audio/eq-fit's Levenberg-Marquardt fit.

Kirkeby, O. & Nelson, P.A. — Digital Filter Design for Inversion Problems in Sound Reproduction, J. Audio Eng. Soc. 47(7/8):583–595, 1999. Oppenheim, A.V. & Schafer, R.W. — Discrete-Time Signal Processing, 3rd ed., §12 (Homomorphic Filtering) — minimum-phase reconstruction from a magnitude spectrum via real-cepstrum folding. Precedents: REW (Room EQ Wizard), Dirac Live, Acourate — commercial/reference room-correction tools this module's shape follows (target curve, smoothing, boost/cut limits, band edges, min-phase option). Olive, S., Welti, T. & McMullin, E. — A Statistical Model that Predicts Listeners' Preference Ratings of In-Room Loudspeaker Response, AES 135th Convention, 2013 — the Harman target curve, one shape for opts.target's {freqs, db} form.

let ir = /* Float32Array impulse response from @audio/measure-ir */
let result = await correct(ir, { fs: 48000, target: 'flat', mode: 'fir' })
// result.coefs — Float64Array, linear-phase FIR, ready for @audio/eq-fir's firEq

import { apply } from '@audio/measure-correct'
await apply(myBuffer, result, { fs: 48000 })   // firEq(coefs) or eq-parametric(bands+preamp)

// parametric — a real EQ's band list, plus an Equalizer APO export
let result = await correct(ir, { fs: 48000, target: 'flat', mode: 'parametric', bands: 8 })
let apo = await toEqualizerApo(result)   // 'Preamp: -5.0 dB\nFilter 1: ON PK Fc 80 Hz Gain -6.0 dB Q 3.82\n...'

Returns { freqs, measured, target, correction, coefs?, bands?, preamp?, residual, rms, maxBoost } — see index.d.ts for the full shape. measured is the smoothed-and-level-aligned curve the correction was computed against; residual is target − (measured + realized correction), predicted from the actual FIR (its own FFT magnitude) or parametric cascade (@audio/eq-fit's response(), preamp excluded — preamp is a broadband headroom offset, not part of the corrected shape, matching eq-fit's own error convention).

correct(), apply(), toEqualizerApo() and toWavIr() are all async: mode: 'parametric' dynamically imports @audio/eq-fit (an optional peer — install it to unlock parametric/both modes; mode: 'fir' never needs it and works standalone), and apply()/toWavIr() dynamically import @audio/eq-fir/@audio/eq-parametric/@audio/encode-wav only for the branch actually used.

Working math happens on an internal log-spaced grid (96 points/octave, ~10 Hz to Nyquist) rather than the linear FFT-bin grid @audio/spectral-target's smooth()/deviation() use — a linear bin grid is too coarse in octave terms at low frequency (at 40 Hz, n=8192, fs=48000: ~4 bins/octave, too coarse for a clean 1/6-oct average) to resolve fractional-octave smoothing down to the bottom of the room-EQ or headphone-EQ band. deviation() is superseded (not reused) for this reason; target − measured and the taper-to-zero-at-the-band-edge shape are carried over from it, and its target() presets are reused directly for named opts.target values.

A deep, narrow (high-Q) dip is not fully flattened by design — regularization exists specifically to prevent the enormous, narrow-band boost that would take, since that boost doesn't generalize across a room's listening positions (matched by REW's and every serious room-correction tool's own advice against fully undoing narrow nulls). maxBoost/maxCut are hard limits regardless; regularization softens how much of the theoretical inverse gets used before those limits even apply.

Use when: designing a room, headphone, or speaker correction EQ from a captured measurement — feeds @audio/eq-fir's firEq/@audio/eq-parametric for real-time application, or toWavIr()'s convolution-EQ WAV for Equalizer APO / Roon / CamillaDSP.

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

MIT © audiojs