How AI cry detection actually works

Cry detection is the feature most people judge a monitor on, and it is described almost everywhere in the vaguest possible terms. "Smart alerts", "AI-powered", and similar phrases skip the thing you actually need to know, which is when your phone will buzz and when it will stay quiet.

This is what happens between a sound in the nursery and a notification on your phone.

The short version

Cry detection listens to the baby unit's microphone and classifies what it hears on the device, using Apple's SoundAnalysis framework. When crying stays above a confidence threshold for about two seconds, the parent unit gets a push notification. The alert clears after roughly five seconds of quiet. Audio is classified on the phone in the nursery and is never uploaded for a server to analyse.

What the baby unit is listening for

The classifier works on a rolling window of about a second and a half of audio, and the windows overlap, so it produces a fresh judgement roughly twice a second. Each judgement is a ranked list of sound categories with a confidence score attached to each.

Three categories are treated as alert-worthy:

• Crying. Infant crying and sobbing.
• Alarms. Smoke and fire alarms, sirens, alarm clocks, car alarms. If the smoke alarm in the hallway goes off while you are downstairs with headphones on, you get told.
• Screaming and shouting. Distinct enough from crying to be worth catching separately, particularly with an older child.

Everything else the classifier recognises is treated as ordinary room sound. Some of it lights up a live indicator on the parent unit showing that there is noise in the room, and none of it sends a notification.

Why it waits about two seconds

A single window of audio is a weak piece of evidence. A cough, a door, a bird outside, a brief squawk on waking: any of these can score high for a fraction of a second and then be gone.

So the detector does not fire on one match. Crying has to stay above the confidence threshold continuously for about two seconds before an alert goes out. That delay is short enough that you would not notice it. It is long enough to discard almost every one-off sound in a house.

The same logic runs in reverse at the other end. Once an alert is active it stays active until the room has been quiet for around five seconds. Without that, a baby pausing for breath between cries would generate a stream of start-and-stop notifications, which is the fastest way to make someone turn alerts off entirely.

Noise filtering, and why it matters more than raw accuracy

A nursery is not quiet. It contains a fan, or an air purifier, or a white noise machine, or a radiator that ticks, or rain against the window. A detector that reacted to all of it would be useless within a night.

The app explicitly treats a long list of sound categories as ambient rather than noteworthy:

• Fans, air conditioning, vacuum cleaners, washing machines, dishwashers, fridges
• Engines, motors, passing traffic
• Wind, rain, thunder, running water
• Hiss, hum, buzz, static, general rumble
• Insects and crickets
• Clock ticks and dripping taps
• White noise

These never raise an alert, and they do not even light the "sound in the room" indicator. If you run white noise as part of the bedtime routine, or through the monitor itself, the detector is built to hear past it. There is more on that setup in white noise and lullabies through the monitor.

The practical consequence is that an alert reaching you carries some information. A monitor that buzzes at the boiler every twenty minutes trains you to stop looking at your phone, which defeats the purpose of having alerts configured in the first place.

What causes false alarms

No classifier is perfect, and being specific about the failure modes is more useful than claiming there are none.

A television or a radio in the next room. Recorded crying sounds like crying. If the baby unit can hear a soap opera through a wall, it will occasionally react to one.

A sibling. Another child crying, shouting or playing loudly in earshot will trigger alerts, because the detector identifies the sound and not the person making it.

Sensitivity set too high. High sensitivity accepts much weaker matches. In a noisy house that means more alerts, and more of them wrong.

The camera parked next to a noise source. A baby unit sitting beside a white noise machine hears mostly white noise. Cries arrive buried under shush, which produces missed alerts more often than false ones. Put the camera on the far side of the cot from the speaker. Where to put a baby monitor goes through distance and placement.

Your own voice during a settle. Going in to comfort a crying child keeps the alert active while you are in the room. That is correct behaviour, and it does surprise people the first time.

If you are getting alerts you do not want, the order of things to try is: move the camera away from any noise source, drop the sensitivity one step, then look at what else is audible from the nursery. Adjusting sensitivity first is tempting and usually addresses the wrong problem.

Sensitivity, and how to set it

Three settings, and they change one thing: how confident the classifier has to be before the two-second timer starts.

• High. Accepts weak matches. Use it in a very quiet house, or if you are worried about missing a grizzle that has not become a proper cry yet. Expect more false alarms.
• Medium. The default, and the right answer for most rooms.
• Low. Only fires on a clear, confident match. Sensible in a noisy flat, on a street with traffic, or if you share a wall with someone.

Pick one and give it three nights before changing it. Sensitivity tuned during a single bad night tends to end up in the wrong place.

What happens when the app is closed

The two devices behave differently here, so it is worth being precise.

The baby unit has to keep running. It stays plugged in, holds the camera and microphone, and does the classifying. Keep the baby app open for video. Locking this phone or switching apps can interrupt camera capture. Keep the device powered and ventilated during monitoring.

The parent unit does not need the app open. Cry alerts arrive as push notifications with the app in the background or fully closed and the phone locked. There is also a Live Activity on the Lock Screen and in the Dynamic Island so the session stays visible without unlocking anything.

If more than one parent device is paired to the same baby unit, each receives its own notifications independently. Neither person has to be watching for the other to be alerted.

Where the classification happens

The audio is analysed on the baby unit itself, in the nursery, by Apple's on-device classifier. Nothing is sent away to be scored. That matters for two reasons. Detection keeps working regardless of what a remote service is doing, and there is no stream of nursery audio arriving at somebody's server for analysis. On-device AI versus cloud AI covers the comparison in more depth.

MonitorUltra also does not record anything. There is no clip history and no audio stored on a server, so there is nothing to retrieve after the fact and nothing to be exposed in a breach.

What cry detection does not do

Worth stating clearly, because the category is full of overclaiming.

It classifies sound. It does not monitor breathing, heart rate or oxygen. It is not a medical device and it does not detect or prevent anything medical. It is a way of hearing your child from another room without staring at a screen, and it does not replace adult supervision or the standard safe-sleep guidance you already follow.

Motion detection is a separate feature working from the camera rather than the microphone, and it has a different set of trade-offs. When to use motion detection and when to turn it off covers those.

New installations of MonitorUltra require an active subscription to start monitoring, including audio. Any introductory trial depends on the selected App Store offer and eligibility. Existing installations with legacy audio access retain that access; it is not an offer for new users.

Web plans cost $7.99/month or $49.99/year, with a 30-day money-back guarantee. You can cancel renewal at any time. App Store prices and trial terms are shown before you subscribe.

Frequently asked questions

How long before I get an alert? Crying has to persist for about two seconds above the confidence threshold, then the notification goes out. In practice you are looking at a few seconds between the sound starting and your phone buzzing.

Does it work through white noise? Usually, because white noise is explicitly on the ignore list. Placement decides how well. Keep the baby unit on the opposite side of the cot from the speaker.

Is the audio sent anywhere to be analysed? No. Classification happens on the baby unit using Apple's SoundAnalysis framework. The live stream to the parent unit is encrypted in transit and relayed, and none of it is recorded.

Will it tell the difference between my baby and a sibling? No. It identifies the type of sound, and has no way to tell one child from another. Another child crying within earshot will raise an alert.

Can I turn cry detection off and keep the video? Yes. Detection is a setting, so you can watch and listen without it, or run it alongside motion alerts, or use it on its own.