What your phone can actually hear at night

Leaving a phone recording sleep at night sounds either useless or slightly alarming depending on your temperament, and both reactions are reasonable until someone tells you what it is actually doing. So here is the whole mechanism, in order, with the limits marked.

What a phone recording sleep at night is doing

There is no sensor in an iPhone that can see sleep. What there is, on a nightstand, is a microphone and an accelerometer, and between them they cover two genuinely different questions.

The accelerometer covers movement. A phone on a mattress picks up the small shifts that travel through the bed frame, and a phone on a nightstand picks up much less but still registers the bigger disturbances. This is the phone version of actigraphy, the movement based approach sleep clinics have used for decades on a wrist worn device. Its strength is timing. Long stretches of stillness look like sleep, clusters of motion look like wake, and the boundaries between them are where the estimate is most trustworthy.

The microphone covers everything movement cannot see. Breathing that gets loud. A cough. A voice. A door. Somebody else’s alarm at five. Those are the events that explain a night, and a movement trace on its own will show you a disturbance without ever telling you what it was.

What the classifier is listening for

The audio side runs through Apple’s SoundAnalysis framework, which is a sound classification system built into iOS. It takes short slices of what the microphone hears and puts a label on them. Tock uses it for the three categories that matter in a bedroom: snoring, coughing, and sleep talking.

That word classification is doing real work in that sentence. The system is not transcribing anything, it is not trying to understand speech, and it has no idea what any words mean. It answers a much smaller question, which is roughly “does this slice sound like the acoustic signature of a snore”. A conversation and a lullaby and someone reciting a phone number are all just sleep talking to it.

The classifier runs on the device. Nothing is being shipped off somewhere to be understood by a server. When Tock’s sleep coach reads your nights and offers to play back what it heard, that part needs a network connection, because the coach is a language model rather than an on device classifier, and that distinction is worth knowing rather than glossing.

Why the clips exist

When something gets classified, Tock saves a short clip of it. Those clips sit in the night detail alongside the rest of the night, and you can play them.

This is the part people underestimate before they have used it. A chart that says you woke at 03:12 is information. A ten second clip of the actual noise that woke you at 03:12 is an answer. Sometimes the answer is a fridge compressor cycling on, or a partner’s phone buzzing against a wooden table, and both of those are fixable in one evening by someone who now knows about them. No chart was ever going to identify either.

It also changes the snoring conversation. Snoring is one of those things people are told about second hand, usually in an argument, and hearing it yourself with a timestamp on it is a very different experience from being told about it over breakfast.

What it is not doing

A few limits, stated plainly, because vagueness here is how apps end up being trusted for things they cannot do.

It is not identifying who made a sound. Two people in a bed produce one audio stream, and nothing in the pipeline separates them. If you both snore, the night shows snoring.

It is not counting breaths as a clinical measurement. Tock shows breathing and room noise across the night as patterns to look at, and respiratory rate as a number only when it comes in from Apple Health, which means from a wearable that measured it properly.

It is not producing measured sleep stages. Movement and sound give an estimate, and the difference between an estimate and a measurement is the whole subject of what estimated sleep stages really mean. Timing is where a phone is strong. Splitting light from deep from REM is where it is guessing well rather than knowing.

It is not always on regardless of what you asked. Tracking runs for the night you started, and both the microphone and the classification are things you grant and can withdraw.

Reading a night with sound in it

The useful habit is to stop looking at the summary number and start looking for pairs. A movement spike on its own is noise in both senses. A movement spike with a clip attached is a story, and a repeating nocturnal awakening that lines up with the same sound three nights running is a thing you can go and fix.

Sound also puts a floor under the estimate. Stillness alone is ambiguous, because lying awake perfectly still and being deeply asleep look nearly identical to an accelerometer. Add audio and the ambiguity narrows, which is a decent argument for a phone on the nightstand over nothing at all. If you want the fuller version of that argument, sleep tracking without a wearable covers what this setup gets right and where a wrist device still wins.

And if the recurring wake is the middle of the night one, why you wake up at 3am is the piece worth reading before you assume the noise caused it. Sometimes the sound woke you. Sometimes you were already surfacing, and the sound is only the thing you happened to notice on the way up.