2026-09-16 · Engineering notes · Anamanti
Good ears on cheap hardware
Getting a repurposed $100 smart display to hear you clearly, over its own voice and in a real room, turned out to be most of the work.
§ Hearing you over its own voice
Cheap smart displays are bad at being interrupted while they’re talking. It turns out the microphone stream on this one carries a time-aligned copy of whatever the speaker is playing. A small shim taps that hidden channel and cancels the echo below Android’s audio system, so “Hey Jarvis, actually tell me about dogs,” said over a playing story, comes through clean. It needs no changes to the app and can be removed cleanly.
§ The bug where the mic lied about its speed
The wake word simply wouldn’t fire, despite clear audio. The device reported its microphone as 48 kHz while actually delivering about 24 kHz, so every sound was processed at double speed, an octave too high, and the model heard gibberish. The fix is to measure the true rate at startup instead of trusting the number the OS reports. I now trust that number about as far as I can throw it.
§ A wake word that runs offline
The “Hey Jarvis” detector is reimplemented from scratch in pure Rust so it runs fully on the device, with nothing leaving the house until you summon it. One side effect of how it’s put together: the expensive part is shared, so adding a second wake word costs almost nothing.