Video Accessibility: Captions, Transcripts and What WCAG Asks For

Who Accessible Video Actually Serves

Accessibility work is often framed as a compliance obligation serving a small minority, which both understates the audience and misses why the work is worth doing. Hearing loss affects a substantial and growing share of the population — the World Health Organization estimates well over 400 million people live with disabling hearing loss — and that figure describes only the permanent case.

The larger group is situational. Someone watching in an open-plan office without headphones, on a train, in a waiting room, or beside a sleeping child has exactly the same need as a permanently deaf viewer for the duration of that session. Someone watching in their second language relies on captions to disambiguate words they can hear but not parse. Someone with an attention difficulty uses captions to stay anchored. The accommodation built for a disability turns out to serve a far broader set of circumstances, which is the usual pattern in accessibility work.

There is also the plain fact that most social video is watched muted by default. A video whose meaning lives entirely in its soundtrack is not communicating to the majority of the people it reaches, which makes captions a distribution decision as much as an accessibility one.

Four Different Accommodations

The terms get used interchangeably, which causes real problems when someone believes an obligation has been met and it has not. There are four distinct things, and they address different barriers.

Subtitles render spoken dialogue as text and assume the viewer can hear everything else. They solve a language barrier, not a hearing one. Captions assume the viewer cannot hear the audio at all, so they carry the dialogue plus everything else in the soundtrack that conveys meaning — speaker identification when it is not visually obvious, relevant non-speech sound such as a knock at the door or a phone ringing, and music cues where the music is doing narrative work. This fuller form is frequently labelled SDH.

A transcript is the full text of the audio as a separate document, untimed. It serves people who prefer reading to watching, anyone using a screen reader, and anyone wanting to search or quote the content — and it is the only one of these that makes a video's contents findable. Audio description is the one most often forgotten: a narration track describing visual information that is not conveyed in the audio, for viewers who cannot see the picture. A slide full of figures that the presenter never reads aloud is inaccessible without it, no matter how good the captions are.

Which WCAG Criteria Apply to Video

The Web Content Accessibility Guidelines organise requirements into success criteria at three conformance levels. For time-based media, the relevant criteria sit in Guideline 1.2, and knowing which ones apply at which level prevents both under-delivering and over-engineering.

At Level A, prerecorded video with audio needs captions (1.2.2), and prerecorded video needs either an audio description or a full text alternative (1.2.3). Prerecorded audio-only content — a podcast — needs a transcript, and prerecorded video-only content needs either a description or an audio track (1.2.1). At Level AA, which is the level most public-sector and procurement requirements actually reference, live video gains a caption requirement (1.2.4) and prerecorded video requires audio description specifically (1.2.5), with the text-alternative escape route no longer sufficient.

In practice, for ordinary prerecorded content, Level AA means captions plus audio description. Most teams deliver captions and stop, which satisfies Level A and leaves 1.2.5 unmet — the most common gap worth knowing about if you are being assessed against AA.

The Accuracy Standard Captions Are Held To

WCAG does not express caption quality as a percentage, but the practical standard applied by broadcasters, universities and accessibility offices is near-verbatim accuracy — commonly stated as 99% or better, including correct punctuation and speaker identification. The reasoning is straightforward: a caption track is a substitute for the audio, and a substitute that is wrong one word in twenty is not equivalent access.

This is the gap between automated output and compliant captions, and it is worth being precise about. A good speech model on clean audio lands somewhere in the mid-nineties, which reads as impressive and sits well below the standard. Worse, the errors are not randomly distributed: they concentrate on proper nouns, technical terms and overlapping speech, which are exactly the words carrying the most information. A transcript that is 95% correct can still misname every person in the recording.

Automated captions also omit, by construction, everything captions are supposed to add beyond dialogue. A speech recogniser transcribes speech; it does not note the doorbell the speaker reacts to, label who is talking, or flag that ominous music has started. Those have to be added by a person who watched the video. The honest framing is that automated transcription removes perhaps eighty per cent of the labour of captioning and none of the responsibility.

A Workflow That Gets You There

The efficient path is to let automation do the mechanical work and spend human attention only where it is actually needed. In practice that is five steps, and the order matters.

Start by generating a draft transcript with timings from the audio. Then do a targeted correction pass: rather than reading everything with equal attention, check names, numbers, technical terms and negations first, because those are both the highest-frequency error sites and the ones where a mistake changes meaning. A missed 'not' is worth more scrutiny than a missed 'the'. Third, restructure the cues for readability — split anything exceeding roughly 17 characters per second or 42 characters per line, breaking at grammatical boundaries rather than mid-phrase, since automated segmentation follows pauses and not sense.

Fourth, add what the model could not know: speaker labels where the voice alone is ambiguous, and bracketed non-speech events that carry meaning. Be selective here — captioning every incidental rustle is as unhelpful as captioning none of it. Finally export both artefacts. The timed file (SRT for editors and platforms, WebVTT for the browser's native track element) serves viewers; the plain transcript serves readers, screen-reader users and search. Publishing the transcript alongside the video costs nothing and is the single cheapest accessibility win available.

• Generate timed draft automatically.
• Correct names, numbers, jargon and negations first.
• Re-split cues for reading speed, breaking at grammatical boundaries.
• Add speaker labels and meaningful sound events by hand.
• Publish both the timed file and a plain transcript.

Testing and Verification

Captions that look right in your editor can fail in delivery, and the failures are usually mechanical rather than editorial. A short verification pass catches almost all of them.

Play the video with captions enabled in the actual player your audience will use, not only in the editing tool. Confirm the track loads at all, that the text is legible against the busiest part of the footage rather than only the title card, and that platform interface elements are not covering it. Check the first and last cues specifically: off-by-one errors at the boundaries are common and easy to miss. Then sample the middle and the end for drift — an offset that is constant throughout indicates a simple timing shift, while one that grows across the runtime indicates a frame-rate mismatch needing a rescale rather than a nudge.

Beyond the mechanical checks, the part that cannot be automated is the part that matters most. Automated checkers verify that a caption track exists and is well-formed; they cannot judge whether it is accurate, whether speaker attribution is correct, or whether a necessary sound event was omitted. Establishing that requires watching the video with the sound off and asking whether you could follow it — and for anything with a genuine compliance obligation, testing with real assistive technology alongside expert accessibility review.

The Side Benefits Nobody Plans For

Accessibility work reliably produces advantages that were not the reason for doing it, and they are worth knowing about because they make the case easier to fund.

The largest is that video contents become text. A video is opaque to a search engine, to your own site search, and to anyone trying to find the moment where a particular thing was said. A published transcript turns the entire recording into indexable, quotable, linkable prose. For a library of recorded talks or lectures, this is frequently the difference between an archive people use and one they do not.

Word-level timings enable interfaces that would otherwise be impractical: click-to-seek transcripts, in-video search, and automatically generated chapter markers. Captions themselves measurably increase watch time on muted autoplay, which is why platforms promote them. And a transcript is the cheapest possible source material for everything downstream — summaries, show notes, blog posts, quotes for social. Having done the captioning work once, several other jobs become nearly free.