I'm working on implementing a similar dictation feature myself, with an emphasis on local models (whisper/parakeet). Very good dictation is hard, and most of the time what makes the difference is a good post processing step, generally using a LLM. This is the only reliable way to remove fillers, cut repetitions, oral idiosyncrasies... I think whisper flow is using a fine tuned version of some llama model. Of course, doing that part locally is more of a challenge because you don't want to have to wait 10 seconds to get the result of your dictation. It is a very interesting challenge.