What Is White Noise? A Complete Sound Guide
A Definition, Stated Precisely
White noise is a random signal whose power spectral density is flat across its bandwidth — equal power in every one-hertz-wide band, from the bottom of the range to the top. That is the whole definition, and it is a statistical statement about a signal rather than a description of a sound. A recording can hiss without being white noise, and a signal can be white noise without sounding the way anyone expects.
The optical analogy that gives the colour its name is roughly true and worth stating once: white light carries all visible wavelengths at approximately equal intensity, and white noise carries all audible frequencies at equal power per unit bandwidth. The analogy stops being useful immediately afterwards, because what follows from the definition is mathematical.
A true white-noise signal would have infinite bandwidth, and therefore infinite total power, which no physical system can produce. Digital audio works with band-limited white noise instead, confined by the sample rate: at 44.1 kHz the signal runs from DC to the Nyquist limit of 22.05 kHz, flat within that range and empty above it. Every real white-noise generator is band-limited, and the term is used loosely to mean the band-limited version.
Why Flat Does Not Sound Flat
Equal energy per hertz is not equal energy per octave, and the gap between those two statements explains most of what listeners notice about white noise. Each octave spans twice the hertz of the one below it. The octave from 10 kHz to 20 kHz contains ten thousand hertz; the octave from 500 Hz to 1 kHz contains five hundred. A flat-per-hertz spectrum therefore delivers twenty times the energy in the first of those octaves as in the second.
Human pitch perception is logarithmic, grouping frequencies into roughly octave-wide bands, so the ear reads that distribution as heavily treble-weighted. The resulting hiss is sometimes described as a flaw in a particular generator. It is not. It is what a flat per-hertz spectrum sounds like to a logarithmic ear, and any correctly generated white noise will have it.
The same arithmetic runs in reverse as a verification method. On a one-third-octave analyser, which divides the spectrum into bands of constant proportional width, white noise should show a rising trend of about three decibels per octave rather than a flat line. A flat display on that instrument means the signal is pink, not white.
The two panels show the same signal, and the contrast between them is the point. Flat in the frequency domain says nothing about the time domain: the waveform is a dense, chaotic band with no visible structure, because successive samples of white noise are uncorrelated. A signal can be perfectly orderly in one domain and look like pure disorder in the other. The small deviations in the left panel are also expected rather than a defect — component tolerances and quantisation keep any real generator from being exactly flat, and holding the deviation inside about a decibel across the audible range is the usual engineering target.
White Noise, Silence, and Ambient Sound
Set three things side by side and the role of white noise becomes clearer. Idealised silence carries no energy at any frequency. White noise carries equal energy at every frequency. Ambient sound sits between them, with energy distributed unevenly according to whatever sources happen to be present and however the room responds to them.
Real environments are never flat. An open-plan office typically has a noise floor dominated by low-frequency energy from ventilation, with intermittent mid-frequency peaks whenever somebody speaks. A bedroom at night often shows a low-frequency rise from distant traffic and building vibration, with very little energy above a couple of kilohertz. Neither profile resembles white noise, and that is exactly why white noise is useful as a reference: it is the one broadband signal with no shape of its own to subtract.
Two properties do the work. Broadband means the signal occupies the full range rather than clustering anywhere. Statistically stationary means its properties do not drift — the mean stays at zero, the variance stays constant, and any segment is statistically like any other segment of the same length. Together these make white noise a dependable test signal, because a measurement taken now and a measurement taken a minute later are measuring the same thing.
What Raising the Floor Does
Perceptually, the consequence is the one most people meet first. In a silent room, every small sound is an event: a dripping tap, a clock, a neighbour's door. Raising the broadband floor reduces how far each of those transients stands above the background, which makes them less likely to capture attention. That mechanism is sound masking, and it depends on the floor being continuous and featureless rather than loud.
Verifying That a Signal Is Actually White
Three independent checks are worth running, because each catches a different kind of failure.
Spectral flatness
A fast Fourier transform decomposes the signal into its frequency components. Averaging many short, windowed segments rather than transforming one long block is the standard approach — Welch's method — and it trades a little frequency resolution for a far steadier estimate. The averaged magnitude spectrum should be flat within the measurement bandwidth.
Autocorrelation
White noise has zero autocorrelation at every non-zero lag, meaning no sample carries information about any other. Finite samples always show small residual correlations, so the test is statistical rather than exact: values beyond lag zero should fall inside the confidence interval for a genuinely random process. This is the check that catches a pseudorandom generator with a short period, which can look spectrally fine while repeating itself audibly.
Crest factor
The ratio of peak to RMS amplitude indicates whether the amplitude distribution is the expected Gaussian. An unusually low crest factor suggests the generator is producing something closer to a uniform distribution, or that the signal is being clipped somewhere in the chain.
A generator can pass one of these and fail another, which is why all three are worth running on a source that other signals will be derived from. Any irregularity in the white-noise source propagates into every filtered variant built on top of it.
Four Things That Are Not True
Several claims about white noise circulate widely and do not survive contact with the definition.
That it is inherently loud. Amplitude and spectral shape are independent. White noise can be generated at any level, and the definition says nothing about volume.
That every hiss is white noise. Tape hiss rolls off at high frequencies because of the magnetic medium, which makes it closer to pink. Radio static between stations carries artefacts of the demodulation circuit that put peaks and valleys in the spectrum. Only a verified flat power spectral density earns the name.
That sample rate is irrelevant. Noise generated at 96 kHz extends to 48 kHz, far above audibility. The ultrasonic content cannot be heard directly, but it can interact with nonlinear processing downstream and change what is heard, so the choice still matters.
That it is unpleasant by nature. The treble emphasis is real and predictable from the spectrum, but whether it is unpleasant is a preference rather than a property. Listeners who find flat-spectrum noise fatiguing generally prefer pink or brown, which is a reason to offer all three rather than a reason to rank them.
Where It Gets Used
Flat spectra make comparisons easy, and most applications of white noise come down to that. Playing it through a loudspeaker and recording the result with a measurement microphone yields the combined response of speaker, room, and microphone; departures from flat are the resonances and nulls. Feeding it through a filter and comparing input to output gives the filter's transfer function without any further analysis.
For sound system calibration the signal of choice is usually pink rather than white, since octave-band analysers assume equal energy per octave, but pink noise is produced by applying a minus-three-decibel-per-octave filter to white, so the white-noise source is still where the chain starts.
In acoustic privacy systems, white or shaped noise raised through ceiling speakers lifts the ambient floor enough to reduce the intelligibility of conversation at a distance. The design goal there is not loudness but consistency and spectral coverage, filling the gaps through which speech would otherwise carry.
References
- White noise — ASA/ANSI Standard Acoustical Terminology, term 2.37
- Welch, P. (1967). The use of fast Fourier transform for the estimation of power spectra. IEEE Transactions on Audio and Electroacoustics 15(2), 70-73
- Pink noise — ASA/ANSI Standard Acoustical Terminology, term 2.38
- Colored noise: spectral slopes of white, pink and brown noise — IEEE Technology Navigator
Frequently Asked Questions
Why does white noise sound hissy even though all frequencies are equal?
Each higher octave spans twice as many hertz as the one below it. Since white noise has equal energy per hertz, higher octaves contain more total energy, making the signal sound brighter and hissier than the name might suggest.
Is white noise the same as static?
Not necessarily. Static from a television or radio contains artifacts from the demodulation process that create an uneven spectrum. White noise is defined by a flat power spectral density, which static may not have.
How can I verify that a noise signal is truly white?
Run a fast Fourier transform on a sufficiently long sample and check that the averaged magnitude spectrum is flat across the frequency range. You can also check the autocorrelation function, which should be zero at all non-zero lags for true white noise.
Does the sample rate affect how white noise sounds?
Yes. A higher sample rate extends the noise bandwidth beyond audible frequencies. While you cannot hear the ultrasonic content directly, it can affect downstream processing and the overall character of the signal when interacting with nonlinear effects.