I personally don’t like speech-to-text because then you’re broadcasting every word you write to everyone around you.

It might be better with a specially-trained model that uses stenographer-style compression that only you know, so drive-by eavesdroppers can’t translate your text entry on the fly.