How cry detection actually works

Every monitor with cry detection describes it the same way, so the feature name tells you nothing about whether it will work in your house. This is what is happening underneath, and which parts are worth asking about.

What the phone is actually listening for

The microphone produces a continuous stream of audio. That stream gets cut into short chunks — in Bebrise, three seconds at a time, mono, at 16 kHz, which is about the quality of a phone call and far more than enough to tell crying from not-crying.

Each chunk goes through a small classifier: a neural network trained on labelled sound, which takes three seconds of audio and returns how strongly it matches each sound it knows. Crying is one of those. It is the same class of model behind "is that a smoke alarm or a doorbell" features, not something exotic.

The model is small enough to run on the phone itself, which is the part that matters most. No audio needs to be uploaded to be understood, because the thing understanding it is already in the room.

Why a good alert waits

Here is where products actually differ, and it is not a detection-accuracy question at all.

A classifier that fires an alert the first time a chunk scores highly will wake you for a cough, a wail on the television, a cat, a door, and a baby who grumbles once and goes back to sleep. Raising the threshold does not fix it: a real cry and a loud short noise both score highly, because they genuinely do sound alike for three seconds. No threshold separates them, because the difference is not in the sound. It is in how long it lasts.

So the useful question is not "was that a cry" but "has this been going on". Bebrise answers the second one. It looks at the last ninety seconds and asks what fraction of that time was crying. Above about half, it decides the baby is crying and tells you. Below about a fifth, it decides that has stopped.

Those two numbers being different is deliberate. If a single figure decided both, audio hovering right at the line would flip the answer back and forth and send you a stream of contradictory alerts. Requiring more evidence to enter the state than to leave it — hysteresis, in control terms — is what stops the flapping. It also means settling back down is not announced the instant a baby goes quiet.

The result is that a single wail never alerts you, and ninety seconds of building crying always does. That is the right trade for the actual situation: you can hear that your baby made a noise. What you cannot judge from another room is whether it is settling or escalating.

The same idea, applied to waking

Waking up gets judged the same way, over a slightly shorter window — sixty seconds, entering at a lower fraction, because a baby lying awake but content moves less than a crying one and that is the case most worth catching early, before the crying starts.

Stillness has to count as evidence for this to work. Babies spend much of the night in active sleep, twitching and shifting for seconds at a time. If the system only recorded the moments something moved, every night would look like constant activity. Quiet intervals are recorded as observed-and-quiet, not skipped.

False positives and false negatives

Both happen. Anyone claiming otherwise has either not tested it or has tuned it to alert on everything and called the noise thoroughness.

What causes false alerts: sustained noise that genuinely resembles crying for a long time. Another child crying nearby. A television left on in the room. Some music. The ninety-second requirement removes almost all the short ones, which is most of them.

What causes missed ones: a very quiet baby, a phone too far away or facing into a corner, heavy background noise — a fan, an air conditioner, rain — sitting under the cry the whole time. And the failure that is not really detection at all: the app was not running.

That last one is the one to design around. A monitor that has silently stopped looks exactly like a monitor with nothing to report. Whatever product you use, find out how it tells you it has stopped — and test the alert path, not just the detection, before you rely on it.

Getting it right in your room

Rooms differ more than the software does. A few things move the needle much further than the sensitivity slider:

• Distance and direction beat sensitivity. A phone a metre from the crib with nothing between it and the mattress hears far better than a distant phone turned up. If you are getting missed cries, move the phone before you touch a setting.
• Continuous noise raises the floor. A fan, an air conditioner, or a white noise machine sits underneath everything and makes a quiet cry harder to pick out. White noise is common in nurseries and there is nothing wrong with it — just expect to place the phone closer if you use one, and to test with the machine actually running.
• A television in the same room is the worst case. Speech and music can hold a high score for long stretches, which is precisely what the ninety-second rule cannot filter out. If a TV is on in the nursery, expect false alerts.
• Soft furnishings help. Curtains and rugs cut the echo that smears sound in a bare room.
• Give it three or four nights before judging. One unusual night tells you nothing, and the first night is always the one with the rubbish bin lorry.

What it cannot tell you

Why your baby is crying. Hungry, bored, uncomfortable, in pain — these are not distinguishable from the sound by this kind of model, and products claiming to translate cries into needs are making a much bigger claim than the technology supports. Bebrise does not attempt it.

Anything about health. Sound recognition is sound recognition. It is not a measure of breathing, oxygen, or distress, and no amount of classifier accuracy turns it into one.

Whether you should go in. It tells you crying has been going on for a while. What that means for your baby, tonight, is entirely yours to judge.

What to ask about any monitor's cry detection

1. Does the analysis run on the device or on a server? If audio is uploaded to be classified, a live microphone feed from your baby's room is leaving the house — which is a bigger question than the feature.
2. Does it alert on a moment or on a duration? Products that fire instantly are the ones people turn off after a week.
3. Can you adjust sensitivity? Rooms differ enormously.
4. Does it work when the internet is down? On-device detection should. If it does not, it was never on-device.
5. How does it tell you it has stopped? The most important question, and the one least often answered.

In Bebrise the audio never leaves the bedside phone, and neither do the camera frames — the reasoning is on the homepage, and the detail is in the privacy policy. The app is in small-scale home testing and not in the app stores yet; ask us for the install details if you want to try it.