How baby cry detection works (and why it isn't just a noise alarm)

What did baby monitors do before "cry detection"?

They measured loudness. The baby unit compares the sound level with a threshold; above it, the parent unit wakes up, lights the sound-level LEDs or beeps. The threshold is what the manual calls sensitivity, and the wake-on-sound behavior is usually called VOX (voice-activated).

Philips's manual for its Avent SCD831/833/835 video monitors puts it plainly: set high, "you will hear many sounds, including soft background sounds"; set low, "only the louder sounds". VTech's monitors have five settings for the same idea:

From VTech's support article "Sound Alert & VOX Sensitivity", checked September 29, 2026.

Notice the wording: "crying and louder sounds". The unit has no idea what a cry is; it knows that crying is usually louder than babbling. That works in a quiet house and fails exactly as you'd expect: a door slam, a shouting sibling or a cough pass the threshold, a tired, quiet whimper doesn't. Turning sensitivity up catches quieter cries and more noise; turning it down drops both. This trade-off is behind most "baby monitor false alarms" complaints.

How does a sound classifier tell a cry from a dog?

The newer approach doesn't ask "how loud?" but "what is this?". A sound classifier is a machine-learning model trained on many hours of audio that people have labeled. Google's AudioSet, the research dataset most of these models grow from, has over 500 sound classes, one of them "Baby cry, infant cry"; next to it sit "Speech", "Dog", "Television", "Door" and hundreds more.

A model learns what each class sounds like: a cry has a rising and falling pitch, harmonics, a rhythm of breaths; a bark is short and broadband; speech has syllables. Fed a second of audio, it returns a confidence for every class it knows. The two Mamoon uses:

• YAMNet (Google, open source) predicts 521 AudioSet classes from audio framed into 0.96-second windows, using the small MobileNet v1 architecture built for phones. Mamoon runs it on Android.
• Apple's SoundAnalysis framework ships a built-in classifier that "can identify over 300 sounds" since iOS 15, with computations "done locally on-device", so "audio is never sent to the Cloud". Its labels include "baby_crying" next to "baby_laughter", "speech" and "dog_bark". Mamoon uses it on iPhone.

The difference for you: a dog barking at full volume scores high for "dog" and low for "baby cry", while a quiet, hoarse cry can still score as a cry. Someone still has to pick the cut-off: a low one catches the first whimper and some laughs and coughs; a high one waits for an unmistakable wail.

Where does the detection run, and why does it matter?

Three places, with different consequences for privacy, latency and what happens when the internet is down.

The practical test: does the alert still arrive with the router up and the internet line down? With a camera-plus-cloud setup it often doesn't; the Tapo cry detection checklist walks through it for one popular camera. The cloud has an upside too: Owlet can capture a video clip with every cry notification. Mamoon keeps no video clips, on purpose: the morning overview stores one still frame per sound, on your phone only.

How Mamoon listens, second by second

Mamoon runs the classifier on the parent's phone. The sound comes from the baby phone's microphone or from an RTSP camera's audio over your Wi-Fi. This is what happens with a real cry:

You pick how eager it should be with three levels:

• "Only crying": alerts only when the classifier is confident it hears a cry. Fewest false alarms; a quiet fuss may pass.
• "Crying and fussing" (default): also alerts on fussing, the grumbling stage before a full cry.
• "Every sound": alerts on any sound above the room's noise, cry or not. The closest to a classic VOX monitor, useful for a very quiet baby or a newborn you want to hear every time.

How accurate is it?

No one can promise a number that means anything in your nursery: accuracy depends on the microphone, the distance to the crib, the background noise, your baby's voice and the level you set. If a product claims "99% accurate" without saying how and where, treat it as marketing. What you can do is test: play a recording of a crying baby by the crib for 15 seconds and watch the state change, then check the morning overview over the first nights to see what triggered alerts.

What cry detection can't do

• It is not a medical device. No cry detector monitors breathing, heart rate or oxygen. It tells you the baby is making noise.
• It has to hear the sound. A white-noise machine next to the microphone, a closed door, or a camera across a large room can hide a cry from any detector.
• It can be fooled by sounds that resemble crying: a baby laughing hard, a sibling imitating, some cats, a squeaky toy. That is what the frame in the morning overview is for.
• It reports, it doesn't reach you by magic. The alert is an ordinary notification; with the phone muted or in Do Not Disturb, it only shows or vibrates. See alerts on a locked, muted or DND phone.

How to reduce false alarms and missed cries

1. Put the microphone within 3–6 feet (1–2 m) of the crib. Nanit gives the same range for its own device. Closer is better for hearing; keep the device out of the baby's reach. Our camera placement guide has the details.
2. Move white noise away from the microphone. A sound machine between the baby and the mic masks the cry; behind or beside the mic it does far less harm.
3. Start with the default level and adjust from the morning overview. False alarms from fussing or babbling: switch to "Only crying". Missed quiet cries: try "Every sound" for a few nights.
4. Check the alert path once. Ringer on and Mamoon allowed through your Sleep Focus or Do Not Disturb. A perfect detector that can't make a sound helps nobody.

Frequently asked questions