Cheap ESP32 units stream raw audio to a server that classifies it and tells Home Assistant — baby crying, smoke alarm, glass breaking, a dog barking, someone at the door. The smoke-alarm detector also runs on-device, so it keeps working if the server or network goes down.
Every event gets its own detector, its own MQTT topic, its own Home Assistant entity — nothing bundled together where one could get lost.
Classified server-side via YAMNet, published the moment it's confident.
Dual-path: on-device and server-side. Survives a server or network outage.
Pairs well with a "nobody home" automation condition — ships as an example.
Its own detector, its own entity — not bundled with anything else.
Catches it even when nobody's near the actual chime.
Split into two distinct alerts, not folded into one bucket — a scream escalates differently than a shout.
Wake-word pipeline is wired end-to-end; real trained models are the next step.
Entities appear in Home Assistant automatically — no manual YAML.
No audio persisted anywhere outside an explicit opt-in debug mode.
ESPHome on an ATOM Echo streams raw PCM over UDP and runs a Goertzel pattern detector for smoke alarms entirely on-device.
YAMNet + openWakeWord classify the stream and publish events over MQTT with hysteresis, not raw noise.
MQTT Discovery creates the entities. Automations and a dashboard ship ready to import.
Heyra hasn't run its distance calibration pass yet (Phase 4, not started) — so instead of a made-up meter figure, here's the honest picture: three room-relative zones, same categories Heyra's own calibration plan will test against.
Where every event type has been exercised so far — loud, distinct sounds (smoke alarm, scream, shout) are the reliable case today.
Target scenario for Phase 4 — expected to work for louder events, not yet walked and measured.
Hardest case, also a Phase 4 target — depends heavily on the specific door/wall.
Get notified the moment crying starts, anywhere in the house.
Distress-keyword detection for someone who can't reach a button.
Glass-break and smoke detection that don't depend on a subscription.
Doorbell and dog-bark alerts even when you're not in the room.
| Board | Status |
|---|---|
| M5Stack ATOM Echo | Verified, in production use |
See firmware/boards/README.md in the repo for the hardware
contract a new board profile needs to satisfy.
Settings → Add-ons → Add-on Store → ⋮ → Repositories
Add https://github.com/AxeForging/heyra
Install the Heyra Add-on — the detection service, plus a live view of your configured units.
Flash each ATOM Echo straight from your browser — no dev machine, nothing installed on the Add-on.
