Could you make it use Parakeet? That's an offline model that runs very quickly even without a GPU, so you could get much lower latency than using an API.
If you do that, the total pipeline takes too long for the UX to be good (5-10 seconds per transcription instead of <1s). I also had concerns around battery life.
Some day!
It’s free and offline
My only issue with it was that it cut off the words [I'm using] at the beginning and obviously it doesn't enter paragraph breaks. It took about 25 seconds to transcribe all of that on a 10th gen i7 laptop processor.
If they could incorporate combination typing out what was said while you're talking it would be pretty perfect.