Transcribing and timing subtitles

Producing a timed transcript from recorded speech is no longer a human operation at scale. The volume of captioned video passed anything a profession could have covered, and subtitlers now correct machine output far more often than they type from the audio. The result is uneven, and Google's own documentation warns that accents, overlapping speech and background noise degrade it, with the errors landing on the people who depend on captions to follow at all. What remains written by a person is the part that is not transcription: reading speed, where to break a line, what to do with a joke or a regional accent in another language, and audio description composed for someone who cannot see the frame. Someone also has to answer when a caption puts words in a speaker's mouth.