How Our Noise Generator Works

Why generate instead of recording?

• No loops: a recording repeats, and once you notice the seam you cannot unhear it. Generated noise never repeats.
• No compression artifacts: streamed audio is compressed, which smears noise. Generated audio is mathematically clean.
• No data use: after the page loads, an hour of sound costs nothing.
• Full control: every layer, filter and level can be changed while it plays.

White noise: random numbers

Digital audio is a stream of numbers, 44,100 or 48,000 per second, each giving the speaker’s position at that instant. White noise is simply a stream of independent random numbers between -1 and 1. Because each value has no relation to the last, every frequency gets equal power. We generate it in an AudioWorklet, a part of the browser’s Web Audio system that runs on a dedicated audio thread, so it keeps going smoothly even when the page is busy.

Pink noise: the Kellet filter

Pink noise needs power to fall 3 dB per octave, which no single simple filter produces. We use the well-known method published by Paul Kellet: white noise is fed through seven simple filters with carefully chosen settings, and their outputs are summed. The result is accurate to within about half a decibel from 9 Hz to the top of the audible range, good enough for calibrating speakers.

Brown noise: a random walk

Brown noise is made by integration: each new sample is the previous one plus a small random step. Left alone, that random walk would drift off forever, so a tiny leak pulls it back toward zero. That gives the 6 dB per octave slope without the drift. Blue and violet noise are made the opposite way, by differencing: subtracting each sample from the next, which tilts the spectrum upward.

Grey and green noise: equalization

Grey noise is white noise passed through three filters that approximate the inverse of the ear’s equal-loudness contour: more bass, a dip around 3.5 kHz and a little extra top. Green noise is pink noise with a broad lift around 500 Hz and gentle roll-offs at both ends.

Nature sounds: noise plus randomness

Rain, wind, waves, fire and streams are built from the same noises, shaped by random events. Rain adds hundreds of tiny, individually tuned drops per second. Wind sweeps a resonant filter with a slow random gust pattern. Waves open and close a filter on a 7 to 13 second cycle. Fire scatters clusters of millisecond crackles over a breathing roar. Every event is chosen at random as it plays, which is why none of them repeat.

Equal loudness between sounds

Raw noises have very different levels: pure white noise is about three times louder than raw pink noise. We measured the average level (RMS) of each generator over a minute and set them all to the same target, then trimmed the bright colors slightly by ear, because at equal energy white and violet noise sound louder. A gentle limiter at the end stops many layers from distorting when added together.

Seamless downloads

When you download a mix, your browser renders it faster than real time, a few seconds longer than requested. The extra tail is then crossfaded into the beginning with an equal-power curve, so the last sample of the file leads directly into the first. Played on repeat, it never clicks or pauses. The file is created on your device and never uploaded.

Listen while you read

More guides