Why Can't Noise Cancelling Block a Crying Baby?

The same pair of headphones that turns a jet engine into a distant hum will do almost nothing about a baby crying four rows back. I assumed for years this was a me problem — a bad seal, the wrong headphones, something I’d fix by spending more. So I spent more. It still happens. It’s the same limitation in every pair on sale, and the reason turns out to be more interesting than “you should have bought the expensive ones.”

It doesn’t block the sound. It adds more sound.

This was the part I had backwards.

Passive isolation is what you’d guess: foam, a tight seal, physical material between your ear and the world. Stuff in the way. Active noise cancellation isn’t that at all. It doesn’t build a wall — it makes more noise, precisely shaped so that the noise and the anti-noise destroy each other.

The analogy that made it click: imagine a heavy door. One person shoves the left side with fifty pounds of force. At the same instant, someone shoves the right side with exactly fifty pounds. The door doesn’t move. Neither person is blocking the other; the forces just cancel.

Sound is a pressure wave — bands of high pressure and low pressure moving through air. Play a second wave that dips exactly when the first one peaks and the two flatten each other out. In acoustics this is called phase inversion, or destructive interference. The anti-wave is the original noise, upside down: 180 degrees out of phase.

So the silence isn’t an absence. It’s two sounds cancelling, and your headphones are working hard to produce the second one.

The loop, and how fast it runs

The whole cycle runs continuously, in a few millionths of a second. Hold onto that, because it’s where my first explanation fell apart.

Where the idea came from

The origin story is a flight in 1978. Amar Bose — MIT professor, founder of the company — was handed one of the airline’s electronic headsets and found he could barely hear anything over the engines. Passive insulation was never going to fix low-frequency rumble; there isn’t enough foam in the world. He reportedly had the mathematical framework sketched out before the plane landed.

Worth noticing that the problem he was solving was specifically aircraft drone. That’s not incidental. It’s the noise ANC is good at, and nearly fifty years later it’s still the noise ANC is good at.

So why does the engine die and the baby survive?

I had a tidy answer for this. It was wrong, and the way it broke is the actual answer, so it’s worth showing both.

The answer I started with was timing — and it doesn’t survive contact with a datasheet. Cancellation means landing the anti-wave at the right point in the cycle. Land half a cycle late and you don’t cancel anything — you reinforce it, and make things worse. Half a cycle at 100 Hz is 5 milliseconds. Half a cycle at 5 kHz is 0.1 milliseconds — a hundred microseconds. Fifty times less room for error. Obvious, then: the chip can’t keep up with the baby.

Here’s what killed it: the chips are far faster than that. One published design for the codec — the chip that does the listening and the arithmetic — puts its analogue-to-digital path at 3 microseconds, and 6 to 8 microseconds for the full trip in and back out. Analog Devices lists a 5 microsecond group delay — the lag the part adds — on a part sold specifically for noise cancelling headsets — though read the small print and that’s analogue in to analogue out with the processing bypassed, so a real implementation with filters running is slower. The general engineering guidance treats anything under about 50 microseconds as comfortable.

Even so, the gap isn’t close. At 5 kHz the hardware has 100 microseconds to play with and the converters spend single digits of it. You could multiply those figures several times over to account for the processing they leave out and still not run out of room. Timing isn’t the wall. My tidy answer was rubbish.

The real limit is that silence is a place, not a moment.

An ANC system doesn’t cancel a sound in the room. It cancels it at a point — wherever the inner microphone is sitting, the one from the diagram above that listens to what actually reaches your ear. Around that point is a bubble where cancellation still roughly holds, and outside it, nothing. Acousticians call this the zone of quiet, and in a diffuse sound field it measures about a tenth of a wavelength across.

That one number decides everything.

At 100 Hz the quiet zone is around 34 centimetres, bigger than your head. Your eardrum sits well inside it and stays inside it when you move. At 1 kHz it’s 3.4 centimetres, already smaller than your ear. By 5 kHz you’re trying to hold a silent bubble seven millimetres wide over a target that moves every time you shift in your seat — move your head five centimetres and your ear has left the zone completely. This is why the research on pushing cancellation to higher frequencies is largely research on head tracking.

Predictability does still matter — an adaptive filter needs a stable pattern to converge on, and a cry gives it nothing. But it’s the second problem, not the first. Even a perfectly clairvoyant system with a zero-latency chip would still only be able to build a pea-sized patch of quiet at 5 kHz.

A baby’s cry lands almost perfectly in the gap. Its fundamental — the lowest pitch in the sound, the note underneath everything else — averages around 450 Hz — typically 400 to 600 — which is low enough that cancellation has some effect on it. But the dominant energy in a cry sits between 1.5 and 3 kHz, and one measurement of a Sony pair put the point where its cancellation gives up at around 700 Hz. The fundamental is inside what the system can handle. Everything that makes the sound piercing is not.

Worse, 2 to 5 kHz is where human hearing is at its most sensitive — the ear canal resonates there, so it’s the band we hear best of all. A cry puts its energy exactly where ANC is weakest and your ear is sharpest. That isn’t a coincidence. That’s what a cry is for.

What this means when you’re buying

Passive and active are good at opposite things, and the split is cleaner than I expected:

• Sustained, low, droning — planes, trains, air conditioning, server rooms, the hum of a city — is ANC’s job, and it’s genuinely very good at it. Feedforward designs — outside microphone — are specced from about 80 Hz to 2 kHz; feedback designs, using the inside one, from about 50 Hz to 800 Hz. Those are the bands the parts are built for, not where cancellation stops being useful in a real pair on a real head, which arrives earlier — hence the 700 Hz measured above.
• Sudden, high, sharp — voices, keyboards, cutlery, a dog — is the seal’s job. Physical isolation, no battery involved.

Which would explain something that always seemed backwards: cheap in-ears with a proper seal holding their own against expensive over-ears at blocking office chatter. That’s not the ANC failing. Office chatter was never ANC’s problem to solve.

What I got wrong, and what’s still missing

The timing explanation is what I believed starting out, and it’s what most write-ups say, including the one that sent me down this hole. I’m leaving it in above rather than quietly deleting it, because it’s a good example of an explanation that sounds rigorous — real numbers, real physics — and still doesn’t survive contact with a datasheet. Latency isn’t irrelevant; it’s a genuine design constraint and it’s why these parts advertise microseconds. It just isn’t what beats your headphones.

What I still couldn’t find: an end-to-end figure for a finished pair of headphones. The silicon vendors publish their numbers. The headphone brands publish “a faster chip.” What happens between the codec and your ear — the driver, the cup, the acoustics — isn’t in any spec sheet I could get hold of. If you know where those live, I’d like to know.

The short version: your headphones aren’t broken and you didn’t buy the wrong ones. You’re asking a system built for three-metre waves to do something about seven-centimetre ones — and the patch of quiet it can hold is only ever a tenth of whichever wave it’s fighting.

Sources

Where the numbers came from, so you can check them rather than take my word for it:

• Optimizing latency in a TWS audio codec — Dolphin Design, via Design & Reuse. The 3 µs converter path, the 6–8 µs round trip, and the feedforward and feedback frequency bands.
• ADAU1787 low power codec — Analog Devices. The 5 µs group delay, on a part sold for noise cancelling headsets.
• Head tracking extends local active control of broadband sound to higher frequencies — Scientific Reports. The zone of quiet at about a tenth of a wavelength, 3.4 cm at 1 kHz, and what five centimetres of head movement does to it.
• Ultra-broadband local active noise control with remote acoustic sensing — Scientific Reports. Further on the size of the controllable zone.
• Acoustical analysis of pain cries in neonates — the 450 Hz average fundamental and the 400–600 Hz range.
• Screaming babies, noise canceling, and you and How to block out crying baby sounds — the 1.5–3 kHz dominant band, the ~700 Hz measured cut-off, and the point that it’s the seal doing the work, not the electronics.

No manufacturer of finished headphones is cited here, because none of them publish the figure that would matter most.