The same video needs more than one kind of text

A local council posts a recording of a public meeting with automatically generated subtitles. A resident searches for the name of a street and finds nothing. Another resident opens the video without sound and sees speech but no description of a crucial graph being presented. A third wants to quote an exchange but cannot tell who was speaking.

The content is 'captioned' in the narrowest visible sense. The information system is still unfinished.

Captions place spoken words and relevant non-speech sounds in time for viewers who cannot or do not use the audio. Subtitles may translate speech for another language audience. A transcript supports searching, quotation and reading without continuous playback. Audio description conveys important visual information not present in dialogue. Chapters help people locate sections.

The W3C WebVTT specification^{[1]W3C WebVTT specification} defines timed text tracks for captions, subtitles, descriptions, chapters and other media-aligned metadata. The fact that one file format can express several kinds of timed text does not mean they carry identical semantics.

Imagine a ten-minute budget presentation. The chair says, 'As you can see, the largest category rose sharply.' A word-for-word transcript preserves the sentence but not the meaning of the slide. A description should add the relevant visual information, while a citation-grade transcript might associate the speaker, timing and source document.

A search index can lead a reader to the right moment, but a search result that quotes a generated phrase as exact official testimony risks changing the record. Automatic speech recognition should retain confidence and review status. Corrections should be traceable back to an immutable recording.

Translating speech is not replacing words one-for-one. Names, measurements, humor and culturally specific terms often need editorial decisions. The original-language track should remain accessible, and a translated track should be identified as a translation with its own producer or review status.

An inaccurate translated caption can shift the apparent position of a speaker. This matters in legal hearings, public comment, journalism and education. Showing uncertainty and allowing corrections is preferable to claiming perfect machine fluency.

Store the underlying recording, source rights, timed utterances, speaker labels where responsibly verified, caption tracks, transcript versions, chapter points and descriptions as related but separable assets. Each rendition can have its own language, producer, review state and provenance.

For a serious public editorial system, a useful listening or viewing interface is not a decorative play button. It is a way to move between original evidence and accessible interpretations without losing their differences.