asked for a full sweep of local speech models, every parakeet variant I could find and the whisper lineup as the alternative, same for the voice side, kokoro, piper, qwen, whatever else exists. the point was a real fallback architecture, what runs where when the other machine is down, and every consumer wired to the same router exactly once.
the results split cleanly. the low precision parakeet variant still drops quiet audio entirely, returns empty text on the file that matters, while the full precision and the alternate build come back clean. directml on the windows box changes the speed but not the accuracy story. the clean path for the server turned out to be an onnx runtime port, that's what goes in the fallback slot. whisper stays in the cloud slot where it already runs, the local whisper variants are slower without a quality win for english.
voice side, kokoro and piper both bench on plain cpu, kokoro becomes the main choice and piper the fallback, and the voice bridge in home assistant comes along for the ride because every consumer gets the same router.
the architecture that came out of it: the main machine runs the good speech to text and the good voice on cpu, the server falls back to the compressed stt model and piper, and the router picks per consumer. the interesting part is that the short audio path and the long audio path are different decisions, short notes stay local and long audio goes to the box that's faster at long files. after that, everything else just points at the router and stops caring where the audio actually gets processed.