Files
zooqueen 2e43622f09 audio: carry word and segment timings through the transcription plane
ProcessAudio returned a bare string, so the timings the upstream
measured had nowhere to go: the decode knew where every word started
and stopped, this plane read {text, duration} and answered {text}. No
caption anywhere in the estate could be cut on a word boundary, and a
consumer that needed one had to place words by dividing a line's span
by its letters.

The transcript is now a Transcript — text, and the timings that were
asked for. timestamp_granularities is read at the edge in either
spelling (the OpenAI SDKs bracket a multipart array name, hand-rolled
clients do not) and forwarded upstream, so word alignment is bought
only when a caller wants it and segments ride along free. verbose_json
answers with task, duration and the timings; plain json stays exactly
{text}. Both transports render that body from one function, and the
provider that reports no timings returns text alone rather than empty
arrays.
2026-08-13 18:18:59 -07:00
..