The Noise Label Problem

Type “brown noise” into any streaming app and you get thousands of tracks. Add “green noise” and you get thousands more. They cannot all be what the label promises, because a noise color is not a mood. It is a number.

This is one of the few claims in the entire ambient-audio catalog that is fully testable. White, pink, brown, green: each name specifies exactly how a sound’s energy is spread across the frequency range. Run the audio through an analyzer and the shape either matches the name or it does not. There is no room for interpretation, no taste involved. So we did the obvious thing. We measured a batch of popular tracks the way a lab instrument would, and asked each one a single question: does the spectrum actually match the word on the label?

The answers were not flattering.

What the claim actually is

A noise color is a spectral slope: a straight line describing how loudness falls as pitch rises, plotted on a log-log graph of power against frequency. The unit is decibels per octave, the drop in energy each time the frequency doubles.

White noise is flat. Equal energy at every frequency, a slope of 0 decibels per octave, which is why it sounds like a bright, even hiss. Pink noise tilts downward at about -3 decibels per octave, softer on the ears and common in nature. Brown noise is steeper still, falling at roughly -6 decibels per octave, the deep low rumble with the treble rolled almost all the way off. Green sits loosely in the middle, weighted toward the mids, and is the least standardized of the four.

Those numbers are the whole definition. “Brown noise” does not mean dark, or bassy, or soothing. It means a spectrum that follows a -6 decibel per octave line. A track is brown or it is not, and the difference is measurable to a fraction of a decibel. The label is a testable claim, and a testable claim can fail.

What the science actually shows

Here is why the label is worth testing rather than trusting. The research people cite when they recommend these sounds was done with carefully specified noise, characterized by its measured spectrum, not with whatever a playlist happens to call brown.

The clearest results are about attention, and they are modest. A 2024 review found that white or pink noise produced a small positive effect on task performance in listeners with elevated attention difficulties, while tending to work against neurotypical listeners (Nigg et al., 2024). Separately, the broader literature on continuous background sound and sleep describes the stimulus as a defined acoustic signal, reporting the spectrum used rather than a color name (Riedy et al., 2021). More recent work on environmental and continuous noise keeps that practice, tying any measured effect to the specific spectral content of the sound (Basner et al., 2026).

The common thread is not the color. It is that every one of these studies knew, and stated, what its noise actually was. That is exactly what a mislabeled file breaks. If a track sold as brown noise is really something else, it does not reproduce the conditions any of that research tested, and there is no reason to expect the same result. The label is not a decoration on top of the science. It is the link between the file and the study, and when the label is wrong, the link is broken.

What we measured

Our method is deliberately boring, which is the point. We take the audio, estimate its power spectral density with a Welch periodogram in scipy, and fit a straight line to the log-log spectrum between 200 Hz and 8 kHz. The slope of that line is the number that matters. White should land near 0, pink near -3, brown near -6 decibels per octave. No listening, no opinion, just the fit.

Run that across the popular catalog and the labels start coming apart. A clip sold as brown noise can measure near -2 decibels per octave, closer to pink than to brown. It masks a different range of frequencies than a true -6 slope would, simply because more high-frequency energy is left in place. It is not lying about being noise. It is lying about which noise. To an ear it may pass; to the fit it does not.

For our own sessions we remove the guesswork entirely. The audio is synthesized in Python from the mathematical definition of the target color, then measured with the same fit before it goes anywhere. The slope we get is the slope we publish, in every description, because it is the one part of the claim that is not a matter of taste. If we call something brown, the receipt says -6.

The honest limits

A measurement settles what a sound is, not what it will do for you. A perfectly correct label does not promise a benefit. The attention effects in the literature are small, depend on the listener, and sometimes run the wrong way, so a true brown noise is not a guaranteed focus aid and is not a treatment for any condition (Nigg et al., 2024).

The measurement has edges too. We fit between 200 Hz and 8 kHz because that is where the colors are best behaved; below and above that window, real recordings wander, and a single straight line is always a simplification of a more textured spectrum. A slope is the right first question to ask of a track. It is not the last one. What it buys you is honesty: you know what you are listening to before you decide whether it helps.

Verify it yourself

You do not have to trust our numbers, and you should not trust anyone’s, including the app that served you the track. Drop any colored-noise file into a free spectrum analyzer and look at the slope. A true brown noise falls off in a straight line at about -6 decibels per octave; pink at about -3; white sits flat. A lot of files labeled brown measure much flatter than that, and once you have seen it on the plot you cannot unsee it.

The label and the measurement should agree. When they do, you have a sound you can reason about. When they do not, you have learned that the word on the file was decoration, and it took about thirty seconds to find out.

Not medical advice. These sessions support relaxation and general wellness, and are not a treatment for any condition.