Moreover, they capture all the smalltalk perfectly, but tend to plausibly mishear the parts that matter most - terminology, abbreviations, names of people. There's no mystery why this happens, of course, but the only transcript I trust is a transcript I personally checked by listening to the audio and fixing mistakes (doable at 1.75x speed, so not that bad). I catch outrageous mistakes sometimes, mistakes that completely change the meaning of what's being said (up to capturing the opposite of what's being said). So, all in all, even though modern speech2text models are really impressive, I am not sure the utility of _completely_ automated transcribers outweighs their dangers today.

It would be nice if they would at least say __unintelligable__ when they couldn't get things, but alas. I also notice it with YouTube auto generated subtitles and translations. They misrepresent SO much.