Closed captions vs subtitles: the real difference

Television separated the two because the technology was different. Subtitles were burned or transmitted as part of the picture, and were mainly used to translate. Captions were encoded in a data channel that the viewer could turn on with a decoder — closed, because the text was hidden inside the signal until activated. Open captions were the version that was always visible.

The terminology survived the technology. Saying "closed captions" today means a caption track that can be toggled; saying "open captions" means the same content rendered into the picture. Neither term says anything about language, which is why a bilingual Chinese-English caption track is still a caption track.

The practical residue is a legal one. Accessibility rules in many countries name captions specifically, and the requirement is about the content — non-dialogue sound, speaker identification, readability — not about the file format. Delivering an SRT with dialogue only and calling it a caption file does not satisfy those rules.

The first addition is speaker identification. When two voices are similar, when someone speaks off-screen, or when several people take turns quickly, a label tells the viewer who is talking. This matters more in Chinese video than in many languages, because names are often the first thing a viewer needs in order to follow a conversation, and because titles and kinship terms carry social information that a bare line of dialogue drops.

The second addition is meaningful sound. A door closing, a phone ringing, laughter, a piece of music that changes the mood — a hearing viewer absorbs these without effort, and a viewer reading captions needs them written down. The editorial judgement is selection: caption what changes the meaning, not what merely exists.

The third addition is readability discipline. Captions are read at the speed of speech, so a line has to fit in a readable span. Practitioners cap characters per line and lines per cue, and prefer two balanced lines over three ragged ones. For Chinese, the constraint is characters rather than words, and the comfortable cap is lower than most people expect.

Songs are the first. Lyrics are fast, metaphorical, and often the emotional core of a scene, and captioning them fully is impossible at reading speed. The usual compromise is to mark the music and caption only the lines that carry the plot, rather than attempting a full transcription that no viewer can follow.

Names and titles are the second. A caption that labels a speaker as "man" when the scene has established his name is worse than no label, because it tells the viewer that the captioner was not paying attention. Use the name the material uses, and keep the naming convention consistent across the whole file.

Dialect and code-switching are the third. Mandarin video frequently contains a line of Cantonese, a regional expression, or an English technical term. A caption track has to decide whether to transcribe the switch as spoken or to translate it, and the decision should be stated once and followed consistently.

Ask three questions before quoting. First, is a track acceptable, or must the text be visible without any player support? That decides closed versus open. Second, does the requirement name non-dialogue sound and speaker labels, or only dialogue? That decides whether you are editing a subtitle file or building a caption file.

Third, what reading speed does the destination enforce? Platforms that check captions for compliance often measure characters per second and flag cues that exceed a threshold. A file that passes visually can still fail that check, and the fix is editing lines rather than re-timing them.

The answers usually reveal that the client needs one of four deliverables: a subtitle track, a caption track, an open-captioned video, or a caption track plus a burned copy for social. Only the last two involve rendering, and only they are irreversible.

Most social platforms autoplay muted, which makes open captions the only reliable way to deliver text. This is why burned-in captions became standard practice on short video long before accessibility rules reached the platforms: the audience was effectively deaf by default.

The design constraints are different from broadcast captions. Text has to survive a small screen and a square crop, so it is larger, shorter, and positioned inside a safe area. Speaker labels often disappear, non-dialogue sound is usually dropped, and the line count per cue drops to one or two.

The honest trade-off is that an open caption for social is not a compliant caption for accessibility. It is visible, which is a real benefit, but it cannot be switched off, cannot be restyled by the viewer, and is often incomplete. If a client needs compliance, they need the track; the burned copy is a distribution convenience on top of it.

Learning content sits between the two categories. A learner can hear the audio and still needs the text, which makes it subtitle work rather than caption work — but a learner also benefits from captions for sounds the vocabulary does not cover, such as a phone ringing in a drama scene being studied for comprehension.

The layer question is specific to this use case. Chinese characters, character-aligned pinyin, and an optional meaning layer do not fit the two-line caption model: three layers need three vertical slots, and a caption spec that expects one line of dialogue has no vocabulary for that. This is why learning subtitles are usually delivered as ASS or rendered video rather than as a plain caption track.

What we do: the transcript and the layers are the product, and the export decides how they reach the viewer. SRT for a plain track, ASS when pinyin has to sit above the characters, and a rendered MP4 when the destination plays video and nothing else.

Several platforms and most public-sector destinations run an automated check before accepting a caption file. The checks are boring and predictable: characters per second, characters per line, lines per cue, minimum cue duration, and overlaps. A file that reads fine to a human can still be rejected for a cue that runs too fast for two seconds.

The practical response is to leave headroom. Write to a reading speed slightly under the limit rather than exactly at it, because the checker measures the cue as delivered and has no idea that the line was meant to be read twice. Where a line cannot be shortened without losing meaning, split the cue rather than speeding it up.

Two Chinese-specific measurements matter more than people expect. Because characters carry more information per unit than Latin letters, a Chinese caption at the same characters-per-second looks denser than an English one at the same number; keep the Chinese rate lower than the English rate you would allow. And because a two-line Chinese cue can be read faster than a three-line one, prefer two balanced lines even when three shorter lines would technically fit the limit.

Finally, keep the checker’s output when it fails. It names the cue, and a named cue is a five-minute fix. A rejected file with no report turns into an afternoon of guessing which line was too fast.