Captions and course accessibility: what actually matters

Captions are the highest-leverage accessibility work available to anyone publishing recorded lessons. They are also the most commonly botched: auto-generated, never reviewed, wrong on exactly the words that carry the meaning. Getting them right is a small, finite job — perhaps ten minutes per lesson — and it changes who can take your course.

Captions are not only for deaf learners

The accessibility case is the primary one: WCAG 2.1 Level A requires captions for prerecorded audio in synchronised media, and in most public-sector and higher-education contexts that requirement is contractual. But the population that benefits is far wider. Non-native speakers use captions to disambiguate unfamiliar vocabulary. Learners in shared spaces watch without sound. Anyone studying a technical subject uses captions to confirm the spelling of a term before searching for it. Captions also make your video searchable, which is the quiet SEO benefit nobody plans for.

What separates good captions from bad ones

Accuracy is the headline requirement, but it is not evenly distributed. A caption track can be 95% accurate overall and still be useless if the 5% of errors fall on domain terminology — which is exactly where automatic speech recognition fails. Names, acronyms, chemical compounds, library and framework names, and non-English loan words are the systematic weak points.

• Timing: cues appear as the words are spoken, with no more than a few tenths of a second of drift.
• Segmentation: 1–2 lines per cue, roughly 32–42 characters per line, broken at clause boundaries rather than mid-phrase.
• Reading rate: below about 180 words per minute on screen; split long cues rather than flashing them.
• Punctuation and case: full sentence casing, because ALL-CAPS blocks are measurably slower to read.
• Speaker labels when more than one person talks, and bracketed non-speech sound only when it carries meaning.

WebVTT is the format to publish

WebVTT is a plain-text format understood natively by every browser through the HTML video element's track tag, and it is accepted by every major course platform. A cue is a timestamp range and one or two lines of text — that is the whole specification you need. Because it is plain text, it is diffable, translatable and editable in any editor, which matters more than it sounds when you are correcting the same mis-transcribed term across twelve lessons.

A ten-minute review that catches almost everything

Do not re-watch the lesson. Instead, work the transcript with a targeted pass. First, build a glossary of the ten to twenty terms that are specific to the lesson and search for each one in the caption file; recognition errors are overwhelmingly consistent, so a single find-and-replace usually fixes every occurrence. Second, scan for cues longer than about two lines and split them. Third, spot-check the first and last thirty seconds, where speech recognition performs worst because of room tone and trailing sentences. Fourth, play back the two or three passages with numbers or code and confirm them by ear.

Machine transcription changes the job from typing to proofreading. Treat it as a draft that needs a glossary pass, not as a finished deliverable.

Translated captions are a separate decision

A translated caption track is not accessibility work — it is localisation, and it should be judged by different criteria. If you teach in English to a French-speaking cohort, translated captions genuinely help. But machine translation of an already imperfect transcript compounds error, so translate from a corrected source track only, and have a speaker of the target language review any lesson where terminology matters. Coursewright generates captions in the lesson's own language and marks the detected language explicitly, precisely so you know which track is the source of truth.

Build it into production, not into a backlog

The reason so many courses ship without usable captions is that captioning gets scheduled after publication, where it competes with new content and always loses. The fix is structural: make the caption file an output of the same step that produces the rest of the lesson materials, so it arrives with the workbook and the quiz rather than as a separate ticket. When captions come out of the pipeline by default, the only remaining work is the review pass — and a ten-minute review is a task people actually do.