Reverb

RMT Compose ships no reverb assets. The impulse response fed to the ConvolverNode is synthesised in the browser, at runtime, by generateImpulseResponse() in src/player/reverb.js (153 lines: one exported pure function, two clamp helpers, no state).

That is a licensing decision as much as a size one: a recorded IR would drag someone else's licence into an MIT-licensed repo. Rendering our own keeps the relicense clean.

Reverb is on by default (audio.reverb.enabled: true, settings-schema.js:72) at 25% wet.

The function

import { generateImpulseResponse } from './player/reverb.js';

const ir = await generateImpulseResponse(audioContext, {
  roomSize: 0.5,   // [0, 1]     — clamped
  decaySec: 1.8,   // [0.1, 12]  — clamped
  damping: 0.5,    // [0, 1]     — clamped
  sampleRate: 48000  // optional; defaults to audioContext.sampleRate, then 44100
});

Returns a Promise<AudioBuffer> — a stereo buffer, rendered in an OfflineAudioContext(2, length, sampleRate). The audioContext argument is used only for its sample rate.

Throws Error('OfflineAudioContext unavailable') if neither OfflineAudioContext nor webkitOfflineAudioContext exists. AudioGraph._regenIR() catches this, warns, and leaves the convolver with whatever buffer it had.

What gets rendered

The IR is decaySec + 0.12 seconds long — the decay time plus a short pad so the tail is not clipped.

1. Decorrelated stereo noise

Each channel is filled with independent white noise. Independent, not a copy: that is what makes the tail wide rather than a mono blob in the centre.

2. An exponential decay envelope

amplitude(t) = noise · build(t) · e^(−t · ln(1000) / decaySec)

ln(1000) (LN1000 = 6.907755278982137) puts the envelope at roughly −60 dB at decaySec, so the slider means what a reverb slider is supposed to mean.

3. A build-up ramp, not a hard onset

build(t) ramps the noise density in linearly over the early-reflection window, instead of starting at full scale at sample 0. A real room's diffuse field builds up; it does not exist at full strength the instant the impulse fires.

Be honest about what this buys

It takes about 2.5 dB off the first 50 ms of the impulse response, and nothing off the peak. Because convolver.normalize = true, the node rescales by total IR energy — so most of the energy the ramp removes is handed straight back as normalisation gain. Net, the late tail you hear in a musical rest comes out about 1 dB hotter, not quieter.

It is kept because a build-up is the physically right shape, not because it quiets anything. The levers on rest-tail level are decaySec and wet.

4. Early reflections — the "size" cue

Seven discrete taps per channel are added inside the early-reflection window:

erWindowSec = max(0.006, roomSize × 0.08)      // 6 ms floor, up to 80 ms
tap_i       = floor(frac_i × erWindowSec × sampleRate)
data[tap_i] += (1 − frac_i) × 0.5 × ±1          // sign randomized per tap

The tap fractions are irregular and different per channel:

const ER_TAPS_L = [0.043, 0.137, 0.271, 0.409, 0.577, 0.719, 0.907];
const ER_TAPS_R = [0.079, 0.191, 0.323, 0.463, 0.617, 0.787, 0.953];

They are irregular on purpose. An earlier version placed the taps on a uniform (e + k) / 7 grid — which is a comb filter. At the default room size the taps landed 274.3 samples (5.715 ms) apart, putting comb teeth every 175 Hz across the wet path (autocorrelation of the tap spectrum at 175 Hz: 0.977). These offsets bring that to −0.05, i.e. no periodicity.

This was hygiene, not a fix

In the rendered IR the taps sit ~29 dB below the noise bed. The comb was undetectable in the output either way. Do not go looking for the "before" sound; there isn't one.

roomSize does exactly one thing: it sets the width of that window. At the default 0.5 the early reflections span 40 ms.

5. Damping — a lowpass that sweeps down across the tail

Air and surfaces eat the highs, so a tail has to get darker as it decays. A fixed lowpass cannot do that. This one sweeps:

hiCut = 16000 − damping × 10000;                    // 16 kHz … 6 kHz
loCut = Math.max(250, hiCut × Math.pow(0.08, damping));

lp.frequency.setValueAtTime(hiCut, 0);
lp.frequency.exponentialRampToValueAtTime(loCut, durationSec);
lp.Q.value = 0.2;

damping	Sweeps from	…down to
0	16 kHz	16 kHz (no sweep)
0.5 (default)	~11 kHz	~3.1 kHz
1	6 kHz	480 Hz

Note what darkening does to the shares: at the default damping, 52% of the IR's energy sits above 5 kHz (the previous law left 64% there), while the 1–5 kHz share rises from 29% to 39% — because the band above it shrank, not because anything was added.

6. A fixed highpass, so the tail doesn't boom

A gentle highpass at 120 Hz, Q = 0.2, ahead of the lowpass in the chain:

noise source → highpass (120 Hz) → lowpass (damping sweep) → offline destination

Parameters, and which ones cost a render

Setting	Default	Range	Needs a new IR?
audio.reverb.roomSize	0.5	0–1	Yes
audio.reverb.decaySec	1.8	0.1–12	Yes
audio.reverb.damping	0.5	0–1	Yes
audio.reverb.preDelayMs	20	0–200	No — live DelayNode param
audio.reverb.wet	0.25	0–1	No — live gain on reverbReturn
audio.reverb.enabled	true	on/off	No — ramps the sends and return to 0

Re-renders are debounced by 250 ms and token-guarded, so a slider drag renders once at the end and a stale render never wins the swap-in race. That machinery lives in AudioGraph, not here — this function is pure and knows nothing about settings.

An IR is also generated at construction, even when reverb is off, so the first time a user enables it there is no pause and no pop. Offline rendering works fine while the AudioContext is still suspended.

A known limitation you cannot fix

A sustained tone's reverb tail warbles — measured at roughly 5.5 dB standard deviation around the smooth decay, at about 3.4 Hz.

This is Rayleigh statistics, not a bug in the algorithm. The tail of a narrowband input is the IR's spectrum sampled over a narrow band, and every IR with a random spectrum does this — recorded ones included.

Allpass "diffusion" cannot help

An allpass filter leaves |H(f)| unchanged by construction. Adding diffusion stages to the IR will not smooth the warble, and any proposal to do so is founded on a misunderstanding. The only levers on how present the tail is in a musical rest are decaySec and wet — 1.0 / 0.12 measures −19 dB against the shipped 1.8 / 0.25.

Notes for maintainers

• Math.random() for the noise is fine here. This function never runs inside a workflow script, and the IR is not required to be reproducible.
• Nothing in this file reads settings, touches the DOM, or holds state. If you need to test it, call it directly with an offline-capable context and inspect the returned buffer.
• The IR is rendered per parameter change, not per note. A render costs an OfflineAudioContext.startRendering() over decaySec + 0.12 seconds of audio — cheap, but not free at decaySec = 12.

See also

• Audio Graph — the send/return the IR plugs into, and the regen guard
• Audio Engine — why a Stop without a declick fade smears through the convolver
• Settings — the Room / Reverb controls