01
Published task benchmark
94.29% for 1-second snore-segment classification and 88.49% for 200-second breathing-window classification, evaluated against PSG reference annotations across n=80 paired nights.
On-device acoustic breathing-event detection · by SomniAI LLC
Breathing-event detection that runs on the device — and that you can verify yourself.
A 56.4 KB model that detects snoring and breathing events from a phone microphone. On-device, no cloud, audio never leaves the device — every night, in real time. Embeddable as an SDK, built by the person who wrote the algorithm.
LIVE breathing · 03:14 PAUSE · 22s
0:000:30now
red region = detected breathing pause
56.4 KB
INT8 model · 9,416 params
0.064 ms
inference · on-device
n=80
person-nights · 40 participants
88.49%
breathing-window accuracy
Small is the point. No inference servers to run. No per-call cloud cost. No audio leaving the device. It's a file you embed.
Built to drop into
- Earbud / hearable hardware
- Mattress & bed sensing
- Smart bedside & ambient audio devices
- Remote sleep screening & telehealth
- Connected health & care platforms
Example categories — if one of these is what you're building, the breathing layer is a file away.
How you'd use it
You feed it microphone frames. The production pipeline returns candidate acoustic events — labels, timestamps and confidence — running on the device via CoreML or TFLite. No inference backend is required. The pilot measures end-to-end behavior on your hardware and in your acoustic environment.
- On-device — runs on the phone or the chip, not a server you pay per call
- No cloud — the audio never leaves the device
- Every night — designed to run continuously, not a one-off scan
- Real-time — a live event stream, not a next-morning batch report
That last one is an opening, not a headline: a real-time stream is something your product can act on — whatever "act" means for your platform. We provide the detection layer; you own what happens next. ApneaSense is detection only; it doesn't respond, treat, or intervene.
Two ways to check us — pick the one that fits you
RIGOROUS Run it on your own data
Three technical preprints, three MIT-licensed repos. The methodology, per-seed metrics and training code are public — bring your own corpus and confusion matrix. We'd rather you trusted your own numbers than ours.
SEE IT LIVE Watch it on your own breathing
SomniSense — our live consumer app — runs the production on-device detection pipeline every night. It shows how minute-level waveforms, detected regions, audio review and whole-night timelines consume that output. The app is productization evidence; your own-device pilot remains the integration benchmark.
88.49% breathing-window accuracy snore 94.29% accuracy 56.4 KB · 0.064 ms on-device n=80 person-nights
Three proof layers — kept separate
02
Production SDK output
The engine turns streaming audio into candidate event fields used by SomniSense: labels, timestamps, confidence, summaries and quality state. Those output fields are the production contract; they do not all inherit the window classifier's accuracy number.
03
Your-device pilot
Your microphones, rooms, placement and users change the operating conditions. A focused pilot measures system behavior in that environment before either side treats integration performance as established.
Where it's strong, where it isn't
STRONG The reachable layer
- Cheap, on-device, real-time, zero-contact — just a microphone
- Runs every night without a clinic, a wearable, or a cloud bill
- Validated against in-lab & ambulatory PSG (n=80)
HONEST What it isn't
- An acoustic proxy — not airflow or blood-oxygen; PSG and continuous SpO₂ measure the physiological event more directly
- A screening signal, not a diagnosis; not FDA-cleared
- Precision-first — a conservative lower-bound estimate
Start with a small pilot
The default first step isn't a contract — it's a focused pilot: your device, your environment, a few nights. Not a big commitment, not a long procurement. The pilot is where you benchmark it in your context.
If it proves out, then we pick the engagement that fits — an SDK license, an integration, or co-development. No pricing tiers, no packages; we scope it to what you're actually building.
Talk to the person who built it
The email goes to the founder of SomniAI LLC — who designed and implemented the detection algorithm, holds the pending US patent (PAT-001), and authored the three papers and open-source repos behind it. No SDR, no sales funnel.
That means faster answers, a real technical conversation, and a direct line from your engineering lead to the person who can actually change the model.
Tell us what you're building.
We'll send the paper, the code, and scope a small pilot — your device, a few nights.