Sleep Stages from Audio

How stage estimation works

Your breathing changes as you move through the stages of sleep, and you can hear the difference. The app looks at how regular your breathing is and how much it varies, using the same recording it already uses for snore detection and breathing disruption detection. There's nothing extra to set up.

About 15 minutes into a recording, the algorithm works out your personal breathing baseline, meaning your typical rate and regularity. From there it measures the rest of the night against your own baseline instead of a population average, and applies thresholds from published research to sort each period into a stage.

Breathing regularity carries most of the weight, but the algorithm also looks at:

• Breathing rate, which tends to slow down as you go into deeper sleep.
• Snoring patterns, since the snoring it detects through the night feeds the same model.
• Time of night, because sleep follows a predictable arc. Deep Sleep bunches up early in the night, and REM grows toward morning.
• Movement, since big changes in sound point to you moving around, which suggests lighter sleep or a brief awakening.

The result shows up as a hypnogram in your nightly summary.

The four stages and their breathing signatures

Each stage leaves its own fingerprint in your breathing:

• Deep Sleep: steady, regular breathing with low variance. This is the restorative stage, and it has the most metronome-like rhythm of any of them. You get most of it early in the night, especially in the first 90 minutes and the first few hours.
• REM Sleep: irregular breathing with high variance. REM is dream sleep. It tends to start around 40 to 50 minutes after you fall asleep and builds toward morning, which is why cutting a night short costs you more REM than anything else.
• Light Sleep: moderate regularity, somewhere between the other two. Light Sleep is the transition stage and the easiest to wake from.
• Awake: movement, sounds, or breathing that doesn't fit with sleep. The app also flags a long silence as a possible awakening when other signals back it up.

Deep Sleep sits at one end of the regularity scale and REM at the other, so the app wants to see consistently low variance before it calls a period Deep Sleep, and it drops out of Deep Sleep as soon as variance climbs. The sound classification running alongside adds the snore and movement context.

The Silence label

Staging from audio has one built-in limit. It can only classify what it can hear. When your breathing is too soft for the microphone to sort into a stage, and nothing suggests you're awake, the app marks that stretch as Silence instead of guessing. This happens a lot for quiet breathers, or when a fan or air conditioner covers the sound.

Silence counts as restful sleep. It goes toward your total sleep time and earns credit toward the restorative part of your sleep score, so a quiet night still reads as a good one. Making up a stage from no evidence would make the data less trustworthy, so the app labels the uncertainty and leaves it at that.

Respiratory rate and breathing regularity

Two breathing measurements feed the stage model, and both show up in your results:

• Respiratory rate is shown in breaths per minute (bpm), worked out from the time between the breaths the app hears. A typical adult range during sleep is 12 to 20 bpm. The app takes a reading every few minutes and shows your average. It sets a personal baseline after roughly the first 10 to 15 minutes of detected sleep, using the middle of those early readings so one stray value can't throw it off. Readings are kept within a plausible range of roughly 6 to 25 bpm, which filters out junk spikes while still allowing the higher rates you can hit during REM or a brief arousal.
• Breathing regularity is how consistent the rhythm between your breaths is, on a scale from irregular to steady. Long stretches of very steady breathing point toward Deep Sleep. A shift into more irregular breathing can mean REM.

Both are estimates from audio, for your own insight. If your respiratory rate is well outside the normal range, bring it to a doctor instead of trying to diagnose yourself.

Does the app use Apple Watch data?

Yes, when it's there. If a wearable like Apple Watch reports sleep stages for the night, Snore Timeline uses the watch's Deep and REM percentages for your sleep score instead of its own audio estimates. A wearable measures those stages better than audio can, so the app defers to it. Without a wearable, it falls back to the breathing and movement estimates. The Apple Watch page covers what the watch adds.

Reading the hypnogram

The hypnogram in your nightly summary charts how you moved through the stages over the night. Each colored band is 5 minutes in a given stage:

• Orange: Awake periods
• Cyan (light blue): Light Sleep
• Indigo (deep blue): Deep Sleep
• Pink/Cyan: REM Sleep
• Mint (soft green): Silence, shown only on nights it occurs

Zoomed in

Read it left to right and you can count how many full sleep cycles you went through, see when you got the most Deep Sleep, and spot where REM landed. Tap any block and a pill shows that stage's full time range, like "Light · 3:50am – 4:15am", and the timeline moves to that moment so you can check whether your snoring episodes line up with a particular stage. Scrub the timeline afterward and the hypnogram's selection line follows.

On nights where a wearable synced sleep data but you didn't record audio, the timeline strip draws a stage ribbon in the same colors, and the status bar reads out the stage and your vitals under the center line, like "Light · 57 bpm · 13 br/min · SpO2 96%". Recorded nights keep the plain waveform.

Typical stage distribution

The nightly summary also shows what percentage of the night you spent in each stage. A typical night might look something like:

• 45% Light Sleep
• 20% Deep Sleep
• 27% REM Sleep
• 8% Awake

Those percentages shift with age, health, stress, and plenty else, so compare your own nights against each other and don't worry about matching a textbook. For scoring, the app treats Deep Sleep of roughly 13 to 23 percent as a healthy target, with 20 percent or more earning credit toward your score, and REM of about 20 to 25 percent. Landing in those ranges can add points to your sleep score. They're general guidelines and nothing more.

How accurate are the estimates?

The classification is based on published research into how breathing regularity relates to sleep stages. Studies that stage sleep from a single physiological signal, like breathing or heart-rate variability, reach roughly 70% agreement with polysomnography, which is the clinical sleep study standard. So the estimates give you a fair picture of your general sleep architecture, and a better one of your trends over time.

We checked the estimate against PSG-Audio, a public sleep-lab dataset with clinician-scored stages, using 57 overnight recordings. The app's stages matched the lab scoring on about 70% of 30-second epochs, in line with the published figure for audio-only staging on the same recordings. That is an offline check against a dataset, and most patients in it had severe sleep apnea, so results on a typical home night can differ. The app has not been clinically validated.

A few things push accuracy up or down. Longer recordings help, because several hours of sleep gives the app more complete cycles to look at. A noisy room, a bed partner, pets, or sleeping far from the microphone all make your breathing harder to detect, and the estimates get worse with it.

Why no stages were detected

Stage estimation needs audible breathing through the recording. If a night shows no stages, it's usually one of these:

• Breathing too quiet. Your breathing may not reach the microphone at a level it can work with.
• Phone too far away. Breathing fades fast with distance. 1 to 2 feet from your head works well.
• Background noise. Fans, air conditioners, and other steady sounds can cover the breathing the app is listening for.
• Recording too short. The algorithm needs about 15 minutes to set your baseline before it can classify anything.

If the whole timeline came up empty, and it's more than the stages missing, see Troubleshooting.

Research references

The stage classification draws on these papers:

• Aging Effects on Cardiac and Respiratory Dynamics in Healthy Subjects across Sleep Stages (Schumann et al., 2010). Shows that respiratory scaling exponents differ measurably between sleep stages. Deep sleep has the most regular breathing, and REM the most variable. Those differences are behind the regularity thresholds used for each stage.
• Using Spectral Continuity to Extract Breathing Rate from Heart Rate and Its Applications in Sleep Physiology (Orphanides et al., 2024). Confirms that deep sleep (N3) has the most stable and predictable breathing of any stage. This is why the app requires consistently low variance to call a period Deep Sleep, and leaves Deep Sleep when variance rises.
• Sleep Stage Classification from Heart-Rate Variability. Reports roughly 70% agreement with polysomnography when staging sleep from a single physiological signal, which is comparable to the breathing-regularity approach used here.