Captions

Captions carry the whole soundtrack, not just the words. Who is speaking, that a door slammed off screen, that the music turned ominous, that the line was delivered in Portuguese: everything a hearing viewer gets from the audio and would otherwise lose. Subtitles assume you can hear and translate what is said, which is why a subtitle track never says “[glass breaks]” and a caption track has to. The version marketed as SDH, subtitles for the deaf and hard of hearing, is the format splitting the difference: subtitle timing and styling, caption content.

Closed means the text ships as its own track the viewer can switch on, off, restyle, or translate. Open, sometimes called burned-in, means the text is part of the picture and nobody can turn it off, which is why social video defaults to it: most feeds autoplay muted, and a large share of viewers keep captions on whether or not they need them. On the web the closed form is a <track kind="captions"> pointing at a WebVTT file, and the player draws the cues; kind="subtitles" is the neighbouring value for the translation case.

Live captioning is a different job with different tooling. CART is a trained stenographer producing a verbatim transcript in real time, and automatic speech recognition is the cheap approximation, good enough for a search index and routinely not good enough for names, jargon, or crosstalk. Auto-captions left unreviewed are the usual reason a caption track fails the people it was written for, and cleaning them up afterwards is most of the work of captioning at all.

WCAG asks for captions on prerecorded video with audio at level A, and on live audio at level AA. A transcript is a related but separate artefact: unsynchronised, readable on its own, and the only route that serves someone who is both deaf and blind through a braille display. Videos with no meaningful audio need neither, and marking those correctly is part of the job too. When you write the cues, keep them short enough to read at speed, break lines on phrases rather than mid-clause, name the speaker when it changes, and put non-speech sound in brackets so it never reads as dialogue.