I pressed PTT on the walkie-talkie app. Asked the agent to read me a note from the Brain vault. Released the button. Three seconds of silence. Then the answer came back as a chipmunk on amphetamines.
The Opus packets I shipped to Zello had the wrong frame size declared in the codec header. The server accepted them. The iPhone decoded 16 kHz audio as if it were 8 kHz and played it at double speed. A two-sentence reply, in a voice I'd carefully picked to sound like a calm Russian man, came out as a meth-bunny screaming.
I stared at the terminal log for twenty minutes before I figured it out. That was the moment I knew this wasn't a Sunday afternoon project. It took three Saturdays, one Thursday I should have been sleeping, and an ElevenLabs bill that taught me what local TTS is for.
But it works now. Causeway Bay at 18:00, hands full of coffee, PTT held to ask Claude whether a project is unblocked. Wet market at 07:30, asking the Brain vault to read out today's open tasks. MTR platform, quietly thinking out loud at a model that pushes back on the weakest part of the argument. The walkie-talkie is in my pocket; the AI lives inside it.
This post is the build.