The audio feature that actually separates snoring from talking in the dark

Distinguishing a snore from a sentence, or a gasp from a cough, is the core problem in acoustic sleep analysis — and the one that most sleep apps paper over. A naive classifier that fires on "loud breathing in the 200–3000 Hz band" will flag every partner conversation, dog bark, and car horn at 2am as "sleep-disordered breathing." That is why event counts from most apps are unreliable.

This is the signal-processing breakdown of the feature set that SleepTrace actually uses, and why it works on-device without a server.

The time-frequency signature of a real snore

A human snore is not a steady tone. It is an aperiodic burst with a strong low-frequency envelope (roughly 120–450 Hz), modulated by the soft palate, with spectral energy tapering by ~2kHz. A sentence has formant structure — clear peaks at 500, 1500, 2500 Hz corresponding to vowels — and a rhythm tied to phoneme timing.

The discriminator that matters:

• Spectral tilt slope. Snoring energy falls off rapidly above 1kHz; speech energy stays flatter. A single tilt coefficient over the 0.5–4kHz window separates most snore/sentence pairs.
• Zero-crossing density in the high band. Snoring is voiced and periodic at the glottal rate (~15–30 Hz, with harmonics). High-band zero-crossing density is low for snoring and high for sibilant speech ("s", "sh", "f").
• Harmonic structure. Snoring harmonics sit on a single fundamental tied to vocal-fold vibration. Speech has a structured harmonic stack that tracks vowel formants.

Combining the tilt slope + high-band zero-crossing density gives a clean separator that runs in a 40-coefficient mel filterbank and costs ~90% of one CPU core on an iPhone SE for a 2-second window.

The breathing pause signature

A gasp or apnea-related pause is, by definition, the absence of a periodic signal. The trick is distinguishing "quiet because paused" from "quiet because the phone is far away." The feature: a cross-correlation of the current 4-second window against the trailing 30-second median breathing rhythm. A real pause shows a sharp drop in correlation that persists past the window length and is accompanied by a sub-baseline spectral flatness (the body is still trying to breathe, quietly).

Why this works without cloud compute

All of these — spectral tilt, zero-crossing density, harmonic-to-noise ratio — are cheap FFTs on 2-second windows. They map cleanly onto a small neural net (under 2MB) that runs entirely in CoreML on-device. No upload. No server-side feature store keyed off the microphone. No privacy question at all.

The full on-device pipeline and the exact feature coefficients are published on the SleepTrace engineering blog.

For builders: start with the discriminator, not the dataset

Most sleep-audio apps collect thousands of hours of labeled data and then ship a dense transformer that runs hot and uploads raw audio. Start instead with a three-feature linear separator (tilt slope, high-band zero crossings, correlation drop) and a threshold calibrated on-device. It will beat the transformer on battery life, privacy, and honesty — and you will have a product you can ship before the cloud bill arrives.