← all posts site ↗

a glance should arm my voice assistant

wake words are the worst part of voice assistants. they miss, they get drowned out by background noise, and saying the same phrase every time you want to talk gets old fast. so the design i landed on makes gaze the primary trigger and the wake word a fallback: a camera sees you face the node, a radar confirms you're in the room, and the node arms itself. then you just talk, and a timeout ends the conversation when your train of thought is done.

head pose beats pupil tracking at room distance. a face landmark model estimates where your head points, you calibrate once per room by looking straight at the node for two seconds, and that's the whole training set. one node per acoustically separate room, because one microphone cannot cover a bedroom, a closet, and a bathroom through closed doors. the mic pick matters: a two-mic board with an on-device audio processor for echo cancellation and noise suppression, tied into esphome and home assistant. the status light does the talking: dim when idle, bright when armed, amber when thinking, blue when speaking, red when muted.

i keep flip-flopping on cameras everywhere. the bathroom camera was the holdout, but it's a private network and i'll be living alone, so fine. the important part is the camera only publishes a couple of booleans, never video.

one real finding from the session: an api key was sitting in plain text in a systemd service file. rotated. check your own service files.