This is the proven Pip voice layer, ported directly, not rebuilt from theory. Native browser speech only — no models to download, nothing that can hang. This is the reliable baseline everything else layers on top of, not a fallback for something better.
idleVoice: detecting…
Transcript will appear here.
Live conversation mode listens, sends what you say to a stub reply function (echoes it back with a small twist, standing in for Cortex's real brain), speaks the reply, waits ~900ms, then listens again automatically — the same turn-taking loop from the Pip breakdown, running unmodified.