It is hard to do right, for exactly that reason. The creators, (called stenocaptionists) use a special keyboard attached to computer software. (The same skills are used for court reporting.)

The operator types in sounds, which are expanded to words by the software, similar to T9 textphone abbreviation. In order to keep up, the operator tries to key in the sound as quickly as possible, not waiting for the end of the sentence.

The worst case is names: I've seen Isaac Asimov rendered "I sack him off". Unfortunately some captionists get around the names issue by never trying, so one sees a lot of "he" and "she" (and those are, of course, best guesses and often wrong).