On-device hotword detection

Hands full, hands tied,
hands frozen. Say it.

Voice wake-word SOS gives a worker a hands-free escape hatch. Say “help help help” — three times, on purpose, so it can't trigger accidentally on the telly — and a 5-second cancellable countdown starts. Audio never leaves the device. iOS + Android.

How the hotword model runs

Four things happen, all on the device.

We chose “help help help” because no English-language programme uses that phrase three times in 5 seconds — but a person in distress does, instinctively. The model is trained on the cadence as much as the phoneme set.

STEP 1

Capture

Mic samples a 1.5s rolling buffer at 16 kHz. Buffer never persists — it's circular RAM, overwritten every loop.

STEP 2

Classify

Tiny on-device CNN (~280 KB) scores each frame for the hotword. Runs ~80mA — well below cellular radio.

STEP 3

Count to 3

Three valid hotword hits within a 5-second window — that's the threshold. Single accidental shout is ignored.

STEP 4

Fire (with cancel)

5-second cancellable countdown begins — phone speaks “SOS firing in 5… 4… cancel?” — tap to abort if false alarm.

Why on-device

Three reasons it runs on the phone, not the cloud.

Privacy

Audio never leaves the device. We don't stream your kitchen, your job site, your patient's living room to anyone's server. The model only fires on the cadence — no transcription, no recording.

Battery

An on-device CNN consumes ~80mA in steady-state listening — less than the cellular radio polling for a tower. A typical shift adds ~3% battery drain. Cloud streaming would add 18%.

Latency & offline

Works underground, in a Faraday-cage warehouse, on a coastal cliff with one bar of GSM. The SOS detection itself doesn't need data — only the resulting alert does, and that fallback chain is what SMS-when-no-data is for.

Sectors where it shines

Where hands aren't available, voice is.

Forestry

Hands on a chainsaw, gloves on. A wake-word is the only practical SOS trigger when both hands are committed.

Elder care

Caregivers in a client's home can't always reach for a button. Voice keeps the trigger discreet — no panic-button visible to a confused or aggressive patient.

Drive / fleet

At the wheel, hands stay on it. Voice trigger keeps the driver compliant with hands-free law while still giving them a one-second escape hatch.

Workers can also configure a personal alternate hotword (e.g. “Yankee Yankee Yankee”) if the default phrase is impractical for their environment — a hospital ward where staff routinely shout “help” for non-emergencies, for example. Configurable per worker from /worker/settings.

Say it into the demo — no consequences.

The demo tenant fires the dispatch chain into a sandboxed channel. You'll see Howie's brief land in the dispatcher view in under 4 seconds.