Horror audio with zero audio files: Web Audio in a three.js game

Lullwood, our free browser horror game, ships no audio files. Every sound in it is generated at runtime with the Web Audio API: looped noise buffers, oscillators, filters, gain envelopes and a stereo panner. A search of all 778 tracked files at commit ce46801 finds no audio file and no audio loader.

We build Lullwood in three.js and Next.js. This post shows what the sound graph is made of, the one audible consequence a player notices (the fog tide’s drone swelling before the fog arrives), and exactly what we searched to say “no audio files”. Line numbers are in engine/forest-engine.js at that commit.

What we searched, and what it found

An absence claim is only as good as the search behind it, so here is the whole search. The commit is ce46801, the tip of release/next when we wrote this.

• Every tracked path, by extension. git ls-tree -r --name-only ce46801 lists 778 files. We matched them against .mp3, .wav, .ogg, .oga, .opus, .m4a, .aac, .flac, .aif, .aiff, .weba, .wma, .amr, .caf, .ac3, .mka, .mp2, .mid and .midi. No match.
• Every file, by content. We extracted the commit and ran file --mime-type over all 778 files, in case audio sat behind a wrong extension. Zero of them are audio/*.
• The code, by reference. git grep -n -E over engine, app, components, lib, shared, public, scripts and services, for decodeAudioData, new Audio(, HTMLAudioElement, <audio, createMediaElementSource and the extensions .mp3, .wav, .ogg, .m4a, .aac, .flac, .opus. Zero matches. The same search shape finds 19 uses of createBufferSource in the engine, so the search itself works.
• The whole tree for the headline word. decodeAudioData appears once in all 778 files, in a design note that says there is none. package.json lists no audio library.

Three video files exist, and they are the nearest thing to an exception. Two are QA recordings under QA_REGRESSION/, outside public/. The third is public/death.mp4, the death sequence, which the page loads in a <video> tag marked muted. None of the three has an audio stream in its container. A search like this shows what the repository contains. It cannot show what a browser extension or a CDN adds, and we make no claim about either.

Noise is the raw material

The function noise() (L4145) fills a buffer with Math.random(), either raw white noise or a leaky running sum of it that works as brown noise. Three looped beds are built from it in startAudio() (L4157), each a three-second buffer with loop = true:

• Wind (L4196): brown noise through a lowpass filter at 340 Hz.
• Insects (L4202): white noise through a bandpass filter at 4,800 Hz. Its gain comes from the time-of-day table, and that gain is 0 at night.
• Rain (L4208): noise through a 2,000 Hz bandpass, held at a gain of 0.0001 until the rainfall event raises it. The comment above it says white noise. The call passes true, which is brown. The sound is what the code does, not what the comment says.

The reverb is also noise. impulse(ctx, 2.4, 3.0) (L4152) builds a two-channel, 2.4-second buffer of random samples multiplied by a falling curve, and a ConvolverNode (L4192) uses it as the forest’s echo. Almost every source feeds both the master gain and that convolver. One thing differs from the map. The terrain comes from a seeded generator and is the same for the same seed. The noise comes from Math.random(), so no two sessions share a wind.

The wind answers your movement. Each frame, in the calm bed, its gain is 0.05 + move01 * 0.10, before a time-of-day factor (1 at night, 0.6 at noon) and a fog-tide duck, and its filter cutoff is 320 + move01 * 900 Hz (L9136-9137), where move01 runs from 0 standing still to 1 at full sprint. A sprint triples the wind’s base gain and opens its filter to 1,220 Hz.

The calm bed is a chord, the chase is a different one

The low drone (L4214-4219) is three sine oscillators at 55, 82.5 and 110 Hz, each detuned by up to 6 cents. Those are a root, the fifth above it and the octave. A 0.05 Hz oscillator, one cycle every 20 seconds, nudges its gain by 0.02 around a base of 0.05.

The chase cue (L4222-4231) is built from the same pitches with one change. Three sawtooth waves sit at 55, 77.78 and 110 Hz, through a 560 Hz lowpass. The fifth is gone, and 77.78 is 55 multiplied by the square root of 2: a tritone, the interval a code comment calls “tritone-ish”. A 110 Hz sine throbs under it, its gain driven by a second oscillator that starts at 2.7 Hz, and a quiet 1,245 Hz triangle sits on top.

The cue is silent (gain 0.0001) until a predator is chasing or bee-lining at you. Then the engine pulls the calm bed out entirely, setting the wind and drone targets to 0.0001 (L9126-9128), and the chase cue comes in at a gain between 0.5 and 1.0. The throb rate climbs from 2.3 to as high as 5.5 beats a second (L9122) with how long the chase has lasted and how close the nearest animal is. Lose it, and the cue falls back to silence while the bed returns.

The fog-tide warning is a gain target, not a cue file

This is the one a player notices. In the fog tide post the clock is a 90-second loop whose telegraph starts 10 seconds before the fog does. Here is how that becomes sound, with no file in sight.

• eventCycleBuildAmount() (eventScheduler.ts L48-53) returns 0 through the calm phase, rises linearly to 1 across the 10-second lead-in, and stays at 1.
• The engine does not use that signal directly. Every frame it eases a second value toward it: fogTideBuild moves a fraction dt / 2 of the way (L8457), a two-second lag.
• Also every frame, the calm bed sets the drone’s target gain to 0.05 * fogTideDroneGainMul(...) times the time-of-day factor, with a 0.3-second time constant (L9138). The multiplier is 1 + build * 1.6, so it runs 1 to 2.6.

At night the drone climbs from 0.05 to 0.13 as the lead-in plays out. It is not an automation curve scheduled ahead on the audio timeline. The game recomputes a target every frame from its own clock and the gain node chases it. That has a plain consequence: the clock only runs while you play, and the retargeting sits inside the playing check (L9124), so pausing stops the swell where it was. Change leadIn from 10 to 6 in the config and the warning shortens with it, because no sample was authored to be 10 seconds long.

Three limits are worth knowing. The swell happens only in the calm bed, so a chase mutes it. It is weighted by your distance to a fog patch, so you hear it near one. And it multiplies a base that depends on your local clock. The game reads the hour once, at load (L385): the drone factor is 1 from 20:00 to 03:59, and 0.08 from 11:00 to 13:59. Played at noon, the same warning runs from 0.004 to about 0.010, 8% of the night swing.

Direction is one StereoPanner

The child’s cry is how you find her in fog. childCry() (L2999) makes a sine oscillator at 420 Hz plus up to 90 Hz when you are close, sweeps it up by a quarter and back in 0.6 seconds, and sends it through a StereoPanner. The pan value (L3004-3008) is the child’s offset projected onto your right and forward axes, then right / hypot(right, fwd). Each cry builds its own nodes, starts them, and stops 0.75 seconds later.

The cadence carries distance as well (L9106-9111): a cry every 5.5 seconds when you are 140 units or more away, closing to every 2 seconds beside her, and a peak gain from 0.04 to 0.09.

That pan formula is the sine of the angle off straight ahead, so it only knows left from right. A cry from directly ahead and one from directly behind both pan to the centre. The cost is covered by text, not by audio: with captions on, the same function writes something like “a child crying · far · behind” (L3020-3023). Footsteps are simpler. The engine accumulates distance walked and plays a short noise burst every 1.9 units (L9146-9149), through a bandpass between 150 and 250 Hz, so the footsteps are tied to ground covered, not to an animation.

What it costs

No file means no sound editor. Changing a sound means changing a number and playing the game. Tests cannot listen either, so the engine exposes hooks that read node values instead; qaProbePredatorVolume() (L5085) returns a gain.

It also means Web Audio behaviour becomes your bugs. Chromium puts a GainNode to sleep once the last sound routed through it ends, and a sleeping node’s gain timeline does not advance, so the node’s .gain.value read the old level, so qaProbePredatorVolume() returned stale values and the predator-audio specs failed at random. Players heard the right level. The fix (L4177-4190) is a silent ConstantSourceNode with an offset of 0, connected to that gain node so it never sleeps. On iOS, the context starts suspended whatever the user did, so startAudio() calls resume() inside the gesture (L4161-4166).

Two things we have not measured: how long startAudio() takes to fill its buffers on a phone, and whether the seam of a three-second noise loop is audible. The code does not crossfade it.

The other half of this story, the files Lullwood does not ship for trees, animals and sky, with a measured first load, is in a night forest with one model file. To hear the graph rather than read it, wear headphones and play Lullwood free in the browser, then wait for the drone to rise.