Most of your audience has the sound off

Two different things with one name

• Subtitles assume the viewer can hear but does not understand: they carry the dialogue, often translated.
• Captions assume the viewer cannot hear: they carry the dialogue and the sound that matters: [machine starts], [laughs], the speaker’s name when it is not obvious who is talking.

The distinction is not pedantry. A video whose meaning turns on a sound (a crack, a door, a change in a machine’s note) is unintelligible with subtitles and intelligible with captions, and for a deaf viewer that is the difference between a piece that includes them and one that does not.

Then there is the third group, which is most people: viewers with working ears watching in a queue, in bed, in an office, on a train, with the sound off because the platform started it muted and turning it on is a decision. Captions serve all three at once, which is why they have gone from an accessibility duty to the default.

The constraint is reading speed, not screen space

The mistake beginners make is fitting the words on screen. The real limit is whether they can be read in the time they are there, while the viewer is also watching the picture.

The broadcast conventions exist because of this:

• Two lines maximum. Three lines covers too much picture and forces the eye to travel.
• Around 37 to 42 characters per line. Wider than that and the eye has to sweep, which costs a fixation.
• Roughly 17 characters per second, or about 160 to 180 words a minute, as a ceiling for comfortable reading. Faster is legible and exhausting.
• A minimum of about one second on screen, even for one word, because anything shorter registers as a flash rather than as text.

When the speech is faster than the reading limit, you edit the caption. Condensing speech into readable text is a real skill and it is not cheating; a verbatim caption that cannot be read has served nobody.

Break lines where the sense breaks

This is the single largest quality difference between amateur and professional captions, and it costs nothing.

Bad:

the machine jams on thick denim which is
why we used to turn the work away

Better:

the machine jams on thick denim
which is why we used to turn the work away

Break at clause boundaries: before a conjunction, before a preposition, between a subject and a long predicate. Never split an article from its noun or a name across two lines. The reason is mechanical: reading proceeds by grabbing phrase-sized chunks, and a break in the middle of a chunk forces a re-read of both halves.

The file, and why a sidecar beats burning in

Subtitles live in a plain-text file, usually .srt, which is readable and editable in any text editor:

1
00:00:04,120 --> 00:00:06,800
the machine jams on thick denim

2
00:00:06,800 --> 00:00:09,400
which is why we used to turn the work away

A sidecar file like this can be switched off, translated, indexed by search engines and read by screen readers. Burned-in captions are painted into the picture and cannot be any of those things, but they are the only kind that survives being downloaded and re-uploaded, which is why social platforms are full of them.

The defensible answer for most people is both: burn in for the vertical social cut, supply an .srt everywhere that accepts one.

Automatic captions, and exactly where they fail

Automatic speech recognition is now good enough that starting from it and correcting is faster than typing from scratch. It fails in a predictable pattern, and knowing the pattern makes the correction pass quick:

• Proper nouns: names of people, places, businesses, products. The model has never seen your client’s surname.
• Numbers and units, particularly currencies and measurements.
• Accents outside the training distribution, which in practice means most speakers outside a narrow band of British and American English. Published error rates differ by accent and dialect by a factor of several, and the direction of that difference is consistent enough to plan around.
• Overlapping speech, which it renders as a plausible single sentence that nobody said.

The last one is the dangerous failure, because it is fluent. It does not look like an error.

Free tools for the whole job: YouTube’s automatic captions with its editor, Whisper if you can run it locally, Subtitle Edit on Windows and Aegisub cross-platform for timing, CapCut and DaVinci Resolve’s free version for burning in. None of this needs a subscription.

The limitation

Captions change what your piece is. A viewer reading captions is spending a share of their attention on text, and cannot look at the picture at the same time: the eye cannot be in two places. This means a caption over a shot that requires close looking is a genuine conflict, not a solved problem.

The consequence for the work is upstream: shots that need to be studied should not coincide with dense speech. That is an editing decision made at the structure stage, and it is one of the places where accessibility stops being a checklist item added at the end and becomes something that shapes the piece.