The spoken transcript of one TikTok video, as one item, without cue timings.

The spoken transcript of one public TikTok video, read logged-out. Takes the FULL video address -- https://www.tiktok.com/@handle/video/1234567890123456789 -- and not a bare id, because the page needs the handle in the path. This is the AUDIO, not the caption and not the words burned on the frames: the caption is the text field of tiktok/post and the on-screen words are tiktok/video_screen_text. Three different texts live in the same video and mixing them up would return plausible wrong data. HOW OFTEN IT EXISTS: measured on 2026-10-01 across 10 reachable videos, only 1 carried a subtitle track. A video with no track answers items: [] and IS charged -- the question was answered and the page was downloaded. Expect the empty answer to be the common one. The transcript comes back as ONE item whose text is the whole speech, because the source is a WEBVTT of automatic speech recognition and its cues are arbitrary time slices, not sentences -- one cue in the measured file is "and you need to stop being immature", mid-sentence. The cue TIMINGS are not published: this schema has no offset field, and inventing one per route is how an API starts speaking two languages. title carries the language the source declares (eng-US in the measurement) or null. id, published_at, thumbnail_url and metrics are null: speech has no identifier, no date, no picture and no counter. Reading it costs TWO network calls on our side -- the video page points at a file on another host -- and still 1 credit, because you asked for one thing. Caches for an hour, the same as tiktok/post and tiktok/video_screen_text, so asking for more than one of them on the same video downloads the page once per endpoint.