Get the words, with the times they were said

Text only — no images, so no vision tokens at all. This is the call to make whenever the question is about what is said rather than what is shown; it is the cheapest endpoint in the API by a wide margin.

The uploader's own captions are preferred when they exist, because a human typed the names; otherwise the audio is transcribed. The response says which of the two happened.

Set language when the audio is mostly music, crowd or engine noise: detection misfires there and a wrong language returns fluent nonsense. Every response also reports whether the audio was speech at all, so music scored over a video is not handed back as dialogue.