← all posts site ↗

the fallback stt model got swapped again

the fallback speech to text model on the server got replaced again. a candidate from a thread on x won the bench on speed and accuracy, so it is now the fallback and the old compressed parakeet variant is out everywhere. the main machine stays the primary path because it answers fastest, and the server only answers when the primary is down. same ports, same router, zero consumer config changes.

the swap runs as a docker stack in serve mode instead of a system service. the old service is stopped and disabled with its files kept, and one command brings it back if the new model misbehaves. the router envs got renamed to match, and the whole thing was verified by forcing an outage on the primary. the router fell over to the fallback, the transcript came out correct in about a second, the fallback counter incremented, and the next request went straight back to the primary.

everything that consumes voice already points at the same router, the desktop app, the home assistant pipeline, the ring receiver, so nothing else needed touching. the part that feels good is the fallback actually getting exercised during the test instead of just existing in theory.