White Noise

What the colours mean. They are a description of how the energy is spread across frequency, borrowed from light. White noise has the same power in every hertz, which makes it the brightest and the most effective masker — and, to most people, the harshest, because human hearing is far more sensitive in the upper middle than at either end. Pink noise has the same power in every octave, so it falls by 3 dB each time the frequency doubles; that is the tilt most natural sound has, and it is why pink is the one people leave running for hours. Brown noise falls twice as fast again, 6 dB per octave, which leaves a deep rush with almost nothing up top — like heavy rain heard from inside, or a motorway a mile away.

How this works without any audio files. Nothing here is a recording. Ten seconds of noise is synthesised in your browser the moment you press Play — white noise from a random number generator, pink from a bank of filters that tilts it, brown by integrating it — and then looped for as long as you leave the tab open. That is why the tool is a few kilobytes rather than a few megabytes, why it works with the network switched off, and why the ten seconds you hear are different from the ten seconds anybody else hears.

Why you cannot hear the loop. A ten-second buffer played end to end and then started again clicks once every ten seconds, because the last sample and the first are unrelated — worst of all on brown noise, which is a wandering signal whose value at the end is nowhere near where it began. So the last quarter-second is folded back over the first, which makes the join a continuation rather than a jump. The three plain colours then have nothing left to give the loop away: noise has no features to recognise. Rain does repeat its pattern of droplets every ten seconds, which is long enough that you have to go looking for it.

Rain is the same noise, shaped. It is pink noise passed through a band — rolled off below 70 Hz and above 11 kHz — with a slow drift in level and two layers of drops over it: a wash too dense to count, and a few nearer ones. Both layers matter more than they sound like they should, and so does the balance between them: drops you can count are hail rather than rain. The nearer drops are resonances rather than bursts of noise — a drop plinks because of the air bubble it traps, which rings at a pitch set by its size. Filtered noise alone gets you a hiss that stops, which is a surprisingly poor imitation. How long it rings matters just as much: water is heavily damped and stops almost at once, where anything ringing on for tens of milliseconds stops sounding like water and starts sounding like metal.

How close it gets, and how that was decided. Rain is not a recording, so it will never sound exactly like the real thing; what it does is give the same masking as a plain colour with something less clinical to listen to. How near it lands is not a matter of opinion here, though. Four freely licensed field recordings of real rain were measured — the balance across frequency, how impulsive it is, and how its loudness wobbles from a fifth of a second up to seconds — and this preset is tuned to land inside the range those four recordings span, on every measure they agree about. Several rounds of tuning it by ear had produced something that measured nothing like rain at all.

Which one to use. For masking a specific noise — traffic, a neighbour, an open-plan office — start with white and come down through pink to brown until the thing you are trying not to hear stops surfacing. Speech is mostly between 200 Hz and 4 kHz, so anything with energy in that range will cover a conversation. For sleeping or working, most people end up on pink or brown, because the top end of white noise is tiring over hours. Try all four at the level you would actually use; the differences are much less obvious loud than quiet.

The sleep timer. Set it and the sound fades to silence over the last ten seconds, landing exactly on the minute you chose. The fade matters more than it sounds like it should: noise cut off abruptly is a change, and a change is the thing most likely to wake somebody who has just got to sleep. The countdown is computed from the finishing time rather than counted down, so it stays correct even when the browser has been throttling the tab in the background for an hour — and the fade itself is scheduled on the audio clock, so it lands on time whether or not anything on the page is still running.

A word about level. Noise is unusually easy to leave on too loud, because it never draws attention to itself the way music does — there is no chorus to make you reach for the volume. Sustained exposure above about 85 dB is what causes permanent hearing loss, and eight hours of it is the point at which that becomes likely; a tool you leave running all night is exactly the shape of thing that gets there. Set it just loud enough to cover what you are covering and no louder, and if you are using it with a child, well below that.

What it does not do. There is no equaliser, no way to mix two sounds together, and no recorded ambience — no café, no thunderstorm, no ocean, because none of those can be synthesised convincingly from filtered noise and all of them would mean shipping audio files. The level is stored on this device but is not part of the link, because it describes your speakers rather than the sound.