The language actually spoken (ISO 639-1). When it isn't the one asked for, answer the words you have (or none): if no piece was in the chosen language, the editor says which language it heard instead of making captions.
The words said, in order, with punctuation. Give each start / end (seconds from the start
of this piece) when your service has them — the word-by-word highlight follows them. Without
times (plain strings, or objects without them) the words are timed from the sound itself,
which is close on steady speech but less exact. confidence (0…1), when your service has
one: words under 0.3 are dropped (likely made up), captions averaging under 0.6 are marked
for the person to check.
What your endpoint answers for one piece of sound (DECISIONS #164, #167). It receives a
POSTto<endpoint>?language=<code>(an ISO 639-1 code;hi-Latnfor Hinglish — Hindi in Latin letters, as people type it; orautoto detect) whose body is the sound: 32-bit float samples, little-endian, 16 kHz, mono, at most 31 s (application/octet-stream). Optional query parameters, to pass on to your service as a hint (e.g. Whisper'sprompt):context— the last words said before this piece;vocabulary— names and words that may be said, comma-separated.