The research behind SomniSense.

What we set out to test

The benchmark question is deliberately narrow: how accurately can a phone-microphone model classify labeled 1-second snore segments and labeled 200-second breathing windows in PSG-referenced recordings?

It does not ask whether a phone replaces PSG, and it does not validate nightly BRI, event boundaries, duration, or diagnosis. Everything below describes the two evaluated classification tasks and their limits.

The validation study

Detection performance was measured against the clinical reference for sleep recording — in-lab and ambulatory polysomnography (PSG) — on nights where a phone and the PSG setup recorded the same person, on the same night, at the same time.

• Sample. 80 paired nights across 40 participants — 10 in-laboratory PSG nights and 70 ambulatory PSG nights with a nasal-airflow cannula. Adults, with and without diagnosed sleep-breathing concerns. The cohort underrepresents some groups on purpose-of-honesty grounds — under-18, BMI extremes, certain demographics — and we say so rather than hide it behind a big aggregate number.
• Recording medium. An ordinary bedside smartphone (a mix of iPhone and Android models from 2018 onward), 50–90 cm from the participant's head. Not lab hardware. The point was to test what users actually have.
• Ground truth. The PSG audio was scored by AASM-trained sleep technicians who were blinded to what SomniSense had said about each segment. Blinded scoring is the part that keeps the test from being circular — the scorers don't get to see our answer first.
• Production-output boundary. Event boundaries, duration, type labels, nightly BRI, and personal trends are product outputs. They require output-specific analysis and do not inherit the accuracy of the published window benchmark.

The full number-by-number breakdown — snore-segment metrics, breathing-window metrics, and bootstrap ranges — lives on /accuracy, written for someone who wants to scrutinize each figure.

How the system actually works

SomniSense isn't one model. It's a cascade of two, and the structure is the part I'm most proud of — because it's what makes a real-time sleep model small enough to live on a phone instead of in a data center.

• Stage 1 — a short-window listener. A small CNN looks at one second of audio at a time and asks a single question: was that a snore? It does this every second, all night.
• Stage 2 — a long-window reader. Those per-second answers, together with two simple loudness features, get assembled into a compact 200 × 3 grid — roughly 600 numbers describing a 200-second stretch of your night. A second CNN reads that grid and decides whether the stretch contained a breathing irregularity.

The trick is that intermediate grid. Most audio models carry a heavy spectrogram around — megabytes per window. By compressing each window down to ~600 numbers before the second stage ever sees it, the whole pipeline shrinks to a model under 60 KB that runs in 0.064 ms on an Apple Neural Engine. That's what "on-device, no cloud" actually requires under the hood.

Certain on-device system details — the gating and inference-triggering logic that surrounds the two stages — are covered by a pending U.S. provisional patent application by SomniAI LLC and are not described in the preprints.

The papers

The complete two-stage cascade is written up as one paper — a Research Square preprint with a citable DOI. Underneath it, three companion preprints document each piece in full — published openly on Zenodo, each with its own DOI (categories cs.LG and eess.AS). None have been peer-reviewed — so I'll call them preprints, not papers with a journal stamp. The DOI links are below.

1. Paper A — A Cascaded Two-Stage CNN Pipeline for Smartphone Sleep-Audio Detection

   The two CNN baselines and the cascade structure, validated under a multi-seed bootstrap so the numbers aren't a single lucky split.

   Yang L. (SomniAI LLC) · cs.LG, eess.AS · Zenodo · doi:10.5281/zenodo.20662374

2. Paper C — Coordinate Attention for 1D Audio-Based Sleep Apnea Detection

   A one-dimensional Coordinate-Attention design for the breathing-event stage — a 93% parameter reduction over the plain baseline, studied across multiple seeds.

   Yang L. (SomniAI LLC) · cs.LG, eess.AS · Zenodo · doi:10.5281/zenodo.20663376

3. Paper E — On-Device Compression for Sleep Apnea Detection

   How the model gets from a research-sized network down to the under-60 KB version that ships — quantization-aware training plus structured pruning, without losing accuracy.

   Yang L. (SomniAI LLC) · cs.LG, eess.AS · Zenodo · doi:10.5281/zenodo.20663768

Why preprints, before peer review

Peer review takes months. Waiting would mean giving you nothing concrete to check in the meantime. So the methodology goes out now, openly, as preprints you can read and argue with — with the explicit caveat that they haven't been formally reviewed. When that changes, this page changes, and I'll say what changed.

What this research is — and isn't

The validation reports classification performance on two PSG-referenced acoustic tasks. It does not turn a wellness app into a medical device or extend automatically to every production output.

• Not a diagnosis. SomniSense reports a product-defined Breathing Irregularity Index (BRI) from microphone audio. BRI is not clinical AHI and does not diagnose obstructive sleep apnea.
• Not a replacement for a sleep study. A diagnosis requires a qualified clinician and appropriate clinical testing. Symptoms such as witnessed pauses, gasping, or persistent daytime sleepiness — not an app threshold — are reasons to consider that conversation.
• Not validated for under-18. The cohort was adults only.

Reading this because of a symptom?

The research is the proof layer. If you got here from something you're actually experiencing, these start from the symptom instead of the method:

• → Waking up gasping
• → Tired after eight hours
• → Loud snoring