transcribe media
The words in audio and video, with timing. One ref: a timestamped transcript, sentence-level (word-level with include_words); media under five minutes answers in the call, longer returns a job and a second call after it finishes returns the words. Cached per library item, so a repeat read costs nothing. refs, up to 20: the words already stored on each file, never transcribing; each row carries state: has_words, being_transcribed or held (paused, job_id), none_yet (with the estimate), no_speech or no_audio. timeline_id: transcribes every uploaded clip on the timeline whose words are unknown, which is what captions, get_transcript and remove_words read. Cloud processing.
Authorizations
Headers
Body
sentences (default): each with start and end. lines: one timed line per sentence as one string, half the size, for reading a long file whole. phrases: speech broken at pauses, each with the gap after it, for choosing cut points. states, with refs: each file's transcript state, no words.
sentences, lines, phrases, states 
