Our data & methodology
What it actually takes to read a night from sound alone — with our real numbers, including the unflattering ones.
The numbers
What our training data really looks like
1,045 hours of consented overnight bedroom audio collected for training our sleep-phase model (as of September 2026). Sounds like a lot — until you need ground truth: usable reference labels, derived from wearable staging on nights where both sources aligned cleanly, amounted to 53 hours at that point. That 20:1 gap between "data" and "data you can learn staging from" is where all the hard work lives.
Another honest number: in one of our own recording tables, only 30 of 111 recorded nights counted as real nights (actual sleep onset, full duration). The rest were aborted starts, tests, or phones picked up again. Deciding what a valid night even is came before any machine learning.
When we ported the analysis from iPhone (Core ML) to Android (ONNX), the two engines agreed to within 2.9×10⁻⁶ on identical input — the same model, bit-for-bit behavior, on both platforms.
The method
How a night becomes a report
The iPhone stays on the nightstand and records the night. An on-device model classifies the audio every 30 seconds into a sleep-phase estimate (awake, light, deep, REM) and tags events — snoring, gasping, reduced-breathing patterns, talking, environmental noise — each with a timestamp and a playable clip. Nothing is uploaded for analysis: the entire pipeline runs on the phone, which is why the app works in airplane mode.
The morning report orders the night by what woke you or disturbed you, with the audio as evidence. Recommendations follow sleep-medicine basics (consistent timing, stimulus control from CBT-I, position and alcohol effects on snoring) — written to be actionable, not to alarm.
Being straight
The limits, stated plainly
Phase estimates from audio alone are estimates. Without a body sensor there is no EEG, no heart rate, no blood oxygen — medically robust staging requires a sleep lab, and we say so in the app. What audio does better than any wrist signal is evidence: you can hear the gasp at 3:12 instead of inferring it from a curve. SleepTrace is a wellness app, not a medical device; loud snoring with breathing pauses belongs in front of a doctor, with your recordings as the conversation starter.