Just a few months ago I set-up a serverless gpu service just using whisper.cpp and it detects very well my very-local spanish modisms. I have an active program running on a vps that hits the serverless gpu with the audio when needed because I went with a telegram bot instead of building an app (and paying for dev fees, yikes!).

I wish at some point apple had something like that for in-device STT but so far I'm very happy with the solution.