I have to say that OpenAI's Whisper model is excellent. If you could leverage that somehow I think it would really improve. I run it locally myself on an old PC with 3060 card. This way I can run whisper large which is still speedy on a GPU especially with faster-whisper. Added bonus is the language autodetection which is great because I speak 3 languages regularly.

I think there's even better models now but Whisper still works fine for me. And there's a big ecosystem around it.

I wonder what the wattage difference is between the iPhone STT and Whisper? How many seconds would the iPhone battery last?