@synthlet/lookahead-limiter

A true-peak brickwall limiter for a master bus

Part of Synthlet

The last node before the destination. It looks ahead by a couple of milliseconds, so it can bring the gain down before a peak arrives rather than reacting after it, and it measures peaks on a 4× oversampled reconstruction — so a signal that never exceeds full scale sample-by-sample, but overshoots between samples, is still caught.

Install

npm i @synthlet/lookahead-limiter

Or npm i synthlet for every module at once.

Usage

import {
  registerLookaheadLimiterWorklet,
  LookaheadLimiter,
} from "@synthlet/lookahead-limiter";

const ac = new AudioContext();
await registerLookaheadLimiterWorklet(ac);

const limiter = LookaheadLimiter(ac, {
  threshold: -1, // dBTP ceiling
  release: 168, // ms, 10-90% recovery
  gain: 6, // dB of drive, applied *before* the detector
  lookahead: 2, // ms - construction-time, sizes the delay line
});

source.connect(limiter).connect(ac.destination);

limiter.threshold.value = -3; // an AudioParam: automatable
limiter.latencySamples; // 102 at 48 kHz with 2 ms of lookahead

Parameters

| Parameter | Type | Unit | Range | Default | | ----------- | --------------------- | -------------------- | --------- | ------- | | threshold | AudioParam (a-rate) | dBTP | −24 … 0 | −1 | | release | AudioParam (k-rate) | ms (10-90% recovery) | 10 … 1000 | 168 | | gain | AudioParam (a-rate) | dB (input drive) | −12 … 24 | 0 | | lookahead | construction option | ms | 0.5 … 5 | 2 |

Three things the table cannot carry:

• gain is a drive, not a makeup gain. It is applied before both the detector and the delay line. A makeup gain multiplies the output after the ceiling is enforced, which would destroy the guarantee outright — so there isn't one.
• release is the 10-90% recovery span, not a time constant. A limiter labelled under the "sum of the two time constants" convention takes about 1.7× longer to recover than its number says; this one does not.
• Automating threshold downward takes lookahead samples to reach the smoothed gain, though the arithmetic backstop reads the current value, so the hard bound holds instantly.

Below the threshold the limiter is a bit-exact passthrough, delayed — not "almost unity". A channel that appears mid-stream starts from a zero-filled delay line, so it fades in over latencySamples.

Latency

latencySamples is the lookahead window plus the detector's group delay (6 samples), and latencyTime is the same figure in seconds. Web Audio has no automatic delay compensation, so if any dry signal is mixed in parallel with the limiter it must be delayed by this much by hand:

const dry = new DelayNode(ac, { delayTime: limiter.latencyTime });

Credits

Original, re-derived from published algorithm descriptions:

• Hämäläinen, Smoothing of the Control Signal without Clipped Output in Digital Peak Limiters, DAFx-02, §3.5 — the gain-smoothing structure.
• ITU-R BS.1770-4, Annex 2 — true-peak measurement. The detection here is BS.1770-style 4× true-peak: the interpolator is a 48-tap Hann-windowed sinc, not the ITU reference coefficients, and its accuracy is measured rather than inherited (within ±0.1 dB from 60 Hz to 10 kHz at 48 kHz, ±0.35 dB from there to 20 kHz — the wider bound near Nyquist is the 4× grid, which every BS.1770 4× detector shares).

The release cascade is in neither source; it is derived from five requirements recorded in src/dsp.ts.

See the repository's THIRD-PARTY-LICENSES.md.

License

MIT © danigb