On-device sleep health from a phone microphone — no wearable, no upload, no account.
SomniAI is an independent research and product LLC working on audio-based sleep monitoring that runs entirely on the device. The repositories here are the algorithm side of that work, openly published so what listens to you sleeping isn't a black box.
The current flagship product is SomniSense, an iOS / Android app that turns one night's audio into a private snore + breathing-pause report. → somnisense.top
Production deployment specifics and the audio pre-processing front-end are covered by pending US patents and not redistributed here.
| Repo | What it covers |
|---|---|
audio-sleep-cnn-baselines |
Two CNN baselines for audio-based sleep monitoring — a 2D-CNN snore detector and a 1D-CNN apnea detector, with multi-seed bootstrap CIs on 80 PSG-paired nights across 40 participants. Methodology-focused: intentionally simple architectures + openly released evaluation code. |
ca1d-sleep-apnea |
A 1D Coordinate-Attention architecture for audio-based sleep apnea detection — a 14,001-parameter network that hits 87% accuracy (a 93% parameter reduction over the standard baseline), with full architecture and training code. |
apnea-compression-pipeline |
An on-device compression pipeline: joint quantization-aware training + L1-structured pruning that takes the model down to 9,416 INT8 parameters and CoreML on the Apple Neural Engine at 0.064 ms / inference, without sacrificing accuracy. |
All three accompanying preprints are forthcoming on arXiv (links live here once first uploaded). All code is MIT-licensed for research and reproducibility.
A few specific reasons, in order of importance:
- You should be able to inspect the model that listens to you sleeping. That's a higher bar than most sleep apps clear today. Publishing the architecture, training protocol, and per-seed metrics is the version of "trust us" that's actually verifiable.
- Bootstrap CIs matter on small biomedical datasets. Single-seed numbers in this regime are routinely misleading. We publish 5-seed × 10,000-iteration bootstrap CIs so the next group doesn't waste time chasing seed-noise.
- Reference numbers help downstream work. A common dataset and identical evaluation protocol across the three preprints — anyone testing a new attention design, a new compression technique, or a new front-end can compare against ours.
What we don't publish:
- The 40-participant audio-PSG dataset itself (consent doesn't cover public release).
- The production audio pre-processing front-end and event-triggered inference scheduler — covered by pending US patents.
Wyoming-registered LLC. The longer version of why this exists is on the marketing site: somnisense.top/story.
We're open to a wide range of collaboration:
- Research collaborations — sleep medicine, mobile health, on-device ML, audio-based biomedical signal processing
- SDK / API integration — consumer health platforms, wearable ecosystems, smart-home audio products
- Clinical partnerships — sleep specialists, sleep clinics, healthcare providers, sleep-related medical-device companies
- Academic / industrial co-authorship on the next round of preprints, particularly around cross-cohort validation and language / device generalization
Reach out at service@somnisense.top.
- General contact:
service@somnisense.top - The app: somnisense.top
SomniAI LLC · 30 N Gould St, Sheridan, WY 82801, USA