The Verification Venue · the quiet you can and cannot buy
The Half of the Noise It Can Erase
Noise-cancelling headphones do not cancel noise. They erase the low, steady, boring half of it and leave the half that carries meaning almost untouched, and that split is not a marketing limit. It is forced by the speed of sound. Play the bench below: build a dead-silent null out of two waves, then watch the roughly 1 kHz ceiling fall out of the physics two independent ways.
The trick is destructive interference. A microphone samples the incoming noise, a chip computes its inverse (same amplitude, phase flipped by 180 degrees), and the driver plays that anti-noise so the two waves sum toward zero at your ear. Two equal sine waves that differ in phase by an angle ψ add to a wave of amplitude 2A·|cos(ψ/2)|. At perfect anti-phase that is zero. A hair off it, and the silence roars back. Start the tone and find the null yourself.
The two-wave bench
One tone, and its anti-wave. Hear and see two waves sum toward silence, then watch a few degrees of error undo it. This ideal null is single-point and single-frequency; Panel B is where the real world breaks it.
Cancellation
– dB
how much quieter the sum is
Residual amplitude
–
2A·|cos(ψ/2)| at matched amplitude
Phase error φ
–°
distance from perfect anti-phase
A pure tone standing in for the noise. The null works at any single frequency.
180 degrees is perfect anti-phase. 0 degrees is in phase, which doubles the sound.
Percent of the noise amplitude. Even at perfect phase, a 10% mismatch caps you at 20 dB.
–
the best cancellation this amplitude allows, even with phase perfect
Snap to the null and the sum trace collapses onto the zero line: the two waves erase each other and the tone disappears. Now drag the phase dial just a few degrees. The reduction meter tells the whole honest story of ANC in miniature: at 5.73° of error you still keep 20 dB, at 18.2° only 10 dB, and by 60° you get nothing at all. Cancellation is a razor's edge in both phase and amplitude. Hold that in mind, because everything a real headphone struggles with is a struggle to keep the phase error small.
The ceiling that the speed of sound builds
A knife-edge null is easy at 200 Hz on a bench. In a real ear cup it gets impossible above roughly 1 kHz, and it fails for two independent reasons that happen to bite at the same place. Sweep the frequency and watch both walls close in.
The ANC ceiling explorer
Wall one: the anti-noise has to be computed and emitted before the wave moves on. Wall two: a single source can only quiet a patch about a tenth of a wavelength across. Both shrink with frequency.
Logarithmic, 20 Hz to 10 kHz. Drive it up and watch both walls fail.
Digital group delay <5 µs (Ole Wolff target) plus 5–10 µs acoustic. 30 µs is a sloppy pipeline.
Each design owns a different band. Hybrid is the widest; the peak is 50–500 Hz.
Its energy is shaded onto the plot. Watch which side of the ceiling each sound lands on.
Wavelength λ = c / f
–
at c = 343 m/s
Zone of quiet ≈ λ / 10
–
region a single source can hush
Timing budget (20 dB)
–
Δt = arcsin(10^(−R/20)/2)/(πf)
–
–
Drag the latency floor and the whole ceiling moves: a faster pipeline reaches a little higher, a sloppy one gives up sooner. That is the proof the wall is set by the clock against the speed of sound, not by a brand. But notice the second wall, the shrinking circle, does not care about the latency slider at all. It is pure geometry, and you cannot buy past it with a faster chip. Both walls land at roughly the same frequency, near 1 kHz, which is why that number keeps showing up.
Why the plane goes quiet but the baby does not
Flip on the three real sounds in Panel B and the answer is one picture. The jet's rumble sits squarely
under the ANC curve and vanishes. The office voice and the baby's cry put their meaning-carrying energy
to the right of the crossover, in the band only the passive seal can touch. Here is the honest
version, in four twists the fifty opposite wave cancels it out
explainers skip.
1. The ceiling is physics, not engineering laziness
Two independent walls bite at once. The latency budget collapses to microseconds above 1 kHz: to null 1 kHz to 20 dB you have about 16 µs to compute and emit the inverse, and only about 1.6 µs at 10 kHz, below any real electro-acoustic pipeline. The zone of quiet a single source can create is only about a tenth of a wavelength across, which is 3.4 cm at 1 kHz (about the size of an ear) but a mere 3.4 mm at 10 kHz, far smaller than the gap between the error mic and your eardrum. You cannot buy your way past either with a faster chip alone, because the second one is geometric.
2. Two separate failure modes, usually conflated
High frequency fails even for a perfectly steady tone (latency plus wavelength). Sudden, unpredictable sound fails even at low frequency, because feed-forward ANC has to predict the next instant of the wave from the last one, and a door slam or a rising cry has no predictable next instant. Engine hum loses on neither wall. A baby loses on both.
3. It does not silence voices, it de-basses them
A human voice's fundamental (roughly 85 to 255 Hz) is below the ceiling and does get
cancelled, so speech through good ANC sounds thinner and more distant. But intelligibility rides on
consonants and fricatives at 2 to 4 kHz, which pass straight through. That is exactly why you can still
follow a conversation and still snap awake at a cry. The honest claim is the warmth is removed, the
words remain
, not voices are untouched
.
4. A baby's pitch is not above 1 kHz, so do not say it is
The cry's fundamental (about 400 to 600 Hz) sits inside the ANC band and gets cancelled like any other low tone. What survives is its 2 to 4 kHz salience-carrying harmonics and formants, plus the fact that it is highly non-stationary. Getting this right is the difference between a true page and a plausible-sounding wrong one. The complete answer: ANC is a low-pass eraser with a soft corner near 1 to 1.2 kHz, the passive foam-and-seal is the high-pass eraser, and the sounds engineered to alarm you (speech, cries, alarms) live precisely in the seam between them.
What it erases, and what it lets through
| Erased by ANC | Survives to your ear | |
|---|---|---|
| Which frequencies | Low, below ~1 kHz | High, above ~1–1.2 kHz |
| Which behaviour | Steady, predictable | Sudden, non-stationary |
| Jet / train rumble (30–200 Hz) | Gone (25–40 dB on steady noise) | – |
| A voice | The low warmth (85–255 Hz) | Consonants at 2–4 kHz (the meaning) |
| A baby's cry | The ~400–600 Hz fundamental | Salience at 2–4 kHz, and its unpredictability |
| Owner of the attenuation | Active anti-noise | The passive ear-cup seal and foam |
The check: every number recomputed in front of you
Nothing here is a stored figure. The page recomputes the interference algebra and the timing and wavelength budgets from the same equations the offline verifier uses, live, for your current settings:
The offline gate reproduces all of it from c = 343 m/s and the interference math alone, and re-checks every cited number: node research/how-noise-cancelling-works/verify-how-noise-cancelling-works.mjs.
Free choices and scope. Panel A is a single-point, single-frequency ideal: a real ear and room never see this null, which is the whole point of Panel B. In Panel B, the interference algebra, the timing budget Δt = arcsin(10^(−R/20)/2)/(πf), and the wavelength λ = c/f with the λ/10 zone of quiet are exact, sourced physics. The shape of the achievable-attenuation curve is what is sourced (25–40 dB in the 50–500 Hz core for premium hybrid on steady noise, rolling to ~0 by 1.2–2 kHz, crossover near 1 kHz); the exact interpolation between those anchors, the ear-scale of 3.4 cm, and the passive curve are an illustrative model. The <5 µs digital group delay and the 12 µs total floor are one vendor's design target used to show the budget, not a constant of nature. The roughly 1 kHz ceiling is a soft 1 to 2 kHz corner, design dependent (feed-forward can reach ~2 kHz), not a wall. The dB figure is premium-hybrid on constant noise; cheaper units and non-stationary sound do less. Infant-cry spectra vary by study, age and health; ~400 to 600 Hz is a representative fundamental range.
What is exactly true here, and what is a model
Exactly true (the sourced physics). Two equal sinusoids of amplitude A at phase difference ψ sum to 2A·|cos(ψ/2)|; the residual for a phase error φ from perfect anti-phase is 2A·sin(φ/2), and a 10% amplitude mismatch at perfect phase caps cancellation at −20·log10(0.1) = 20 dB. The reduction ladder −20·log10(2·sin(φ/2)) gives 20 dB at 5.73°, 10 dB at 18.2°, 3 dB at 41.4° and 0 dB at 60°. The timing budget Δt = arcsin(10^(−R/20)/2)/(πf) gives about 159 µs at 100 Hz, 15.9 µs at 1 kHz and 1.59 µs at 10 kHz for a 20 dB target. Wavelength λ = c/f = 343/f is 3.43 m at 100 Hz and 3.43 cm at 10 kHz, so the λ/10 zone of quiet is 3.43 cm at 1 kHz and about 3 mm at 10 kHz. A 12 µs latency floor and the λ/10 geometry independently place the ceiling near 1 to 1.3 kHz. These are reproduced from first principles in the verifier.
A model, not a measurement (the plot). The achievable-attenuation curve and the passive curve are an interpolation pinned to sourced anchors, not a measurement of any specific headphone. It gets the directions right (which frequencies erase, which survive, where the two regimes hand off) without claiming to be a particular product.
Named simplifications. Single anti-noise source and a single error point; a pure tone standing in for broadband noise; no room reflections; a fixed ear-scale for the zone-of-quiet drawing; and the deliberate ~1.2 kHz band-limit is one vendor's choice, not universal. None of these change the reversal at the heart of the page: ANC erases the low, steady half and leaves the meaning.