Hmmm. It looks like it's _possible_ to get voice-to-standard-out working, using consentfully obtained training data, with Vosk.
https://github.com/alphacep/vosk-api/blob/master/python/README.md
I'm going to have to figure this out. "Open weight" models are clearly bullshit, but if a small model I can reproduce out of known-consentful training data can work that would be interesting.