Built from research. Tuned to real recordings. Checked on every change.

1. Built from the research

Every generator starts from how the real thing makes its sound, as the acoustics and physics literature describes it — not from a recording and not from a guess. Some of what that means in practice:

• Water is bubbles. Each one rings at a pitch set by its size, and a river is thousands of them; how many, how big and how they move follows the studies of flowing and falling water.
• Fire breathes at a rate set by its width — a half-metre campfire puffs about twice a second — and its roar, lapping and crackle come from what combustion research says each one is.
• Rain is single drops, each a click and, often, a bubble, on the surface it lands on.
• Wind is what it touches: leaves, grass, wires and edges each answer it in their own way.
• Birds sing through a model of the bird’s own voice box, and answer each other the way studies of birdsong say they do.
• Underwater follows marine bioacoustics: what whales, dolphins and shrimp actually sound like, and how sound travels through water to a hydrophone or a diver’s ears.

Among the work it builds on: Minnaert and van den Doel on bubbles; Tonolla and colleagues on river noise; Cetegen and Ahmed, Chadwick and James on flames; Prosperetti and Oguz on raindrops; de Langre and Moulia on plants in wind; Goller and Mindlin, Amador and Mindlin on birdsong; Madsen, Zimmer, Wisniewska and others on whales and dolphins; Essl and Cook, Schumacher and Woodhouse on bowed sound; McDermott and Simoncelli on how the ear hears textures.

2. Tuned to real recordings

Physics gets the sound into the right place. Recordings show how far it still is from the real thing. So each generator is held against recordings of what it makes — most of them from the BBC Sound Effects archive: for Water, 64 recordings of brooks, slow rivers, lake shores and drips; for Fire, the BBC’s recorded campfire, bonfire and wood fires; for Rain, rain on turf, leaves and roofs; for Underwater, about 110 recordings of reefs, whales, dolphins and surf heard from below; for Electric, 27 of hum, strip lights and welding.

The recordings are never used in the sound. They are measured — how bright, how steady or crackly, how fast it moves — and the plugin is measured the same way. Then the plugin’s own knobs are turned, by a program, until the two match as closely as the knobs allow. Whatever is left over tells us what no setting can fix, and that is the part of the engine we change.

What the recordings taught us

• The hiss of running water fades smoothly toward the top. Every creek, river and waterfall we measured gets steadily quieter above its gurgle; ours had a hump there. It does not any more.
• Slow water is single sounds, not quiet noise. A lake shore or a slow river is a gurgle here, a plop there. Ours was a softer rush. It now has its own sparse, deeper bubbles that take over as the water slows.
• A close wood fire is dark. Its sound sits in the low roar and the lapping; the crackle is well below. Ours was too bright, and its flames moved about twice as much as real ones.
• Rain on grass is softer than you expect — the drops sit lower in pitch than ours first did.
• Not every idea survives. A change to the humpback whale that sounded right made the match worse, so it was taken out again.

3. Checked on every change

Every time the code changes, every plugin is played through a set of fixed situations and measured again — automatically, before anything ships. We call these tests fences: each one marks a line the sound must not cross. There are about 140 of them across the suite, holding thousands of individual checks. A few of them, in plain words:

• Nothing changes by accident. Fixed scenes are played and fingerprinted. If a change was meant to be silent, the sound has to come out identical, down to the last bit.
• Bouncing twice gives the same file. Render a preset twice and you get the same audio. This test caught a delay that, once in a while, read one sample from the wrong place.
• A processor adds nothing of its own. The spatial compressor and gate are fed two steady tones and must not add anything audible. They once did — a faint crackle when pushed — and the test is how it was found and proven gone.
• No knob jumps or clicks. Every knob is swept and automated and the output is checked for steps.
• Presets change without a pop, every time, on a playing sound.
• It runs at 10th order. Each plugin’s processor load on a 121-channel track is measured and must stay within its budget.
• The windows are looked at. Every plugin window is drawn to an image and checked for controls that overlap or fall off the edge.

So yes: the plugins keep being measured. Not once before release, but on every change, for as long as they are developed.

Why it matters

None of this shows in a screenshot. A waterfall that sounds like a waterfall is easy to promise. That it was built from how falling water makes its sound, held against real waterfalls, changed where the measurements said and left alone where they did not, and checked again every time anything changes — that is the work. It is why the sounds hold up when you move them, place them far away, or listen at 10th order, where a sound made by ear tends to come apart.