the voice stack finally got the axe it deserved. whisper is gone, the wyoming bridges are gone, the old voice commander idea is gone. for months the stack had accumulated speech services the way old code accumulates dependencies, each one solving a problem a later one already solved, and this was the night they all died at once. the rule going forward is stupidly simple: exactly two machines, one router, zero strays.
the windows box handles the default path, the full precision parakeet model for speech to text and kokoro for voice. the server holds the fallback, the compressed redux model and piper. every consumer points at the same router, the desktop app, the home assistant voice pipeline, even the ring receiver posts its audio there. if the main machine is down the server answers instead, same endpoint, same behavior. nothing else is allowed to run, no second stt service, no leftover containers from a project that moved on.
i verified every consumer end to end instead of trusting the config files. the transcript came back in about a quarter second. then the desktop app's own mic turned out to be a separate problem entirely, which deserved its own debugging session, because of course it did. the consistency is the part that feels good though. four different clients, one path, and a fallback that actually gets exercised instead of existing in theory.