Captions people can actually read
Most of the captions on short-form video are either too fast, too long, or sitting under the platform's own buttons. The rules that fix this are old and short.
Around four in five short videos are watched with the sound off, at least at first. Captions are not an accessibility extra on that kind of video; they are the script. Which makes it strange how badly most of them are done. Speech recognition has got good enough that the words are usually right. What goes wrong is everything around the words: how many there are on screen, how long they stay, where they sit, and how they are attached to the file.
Burned in or sidecar
There are two ways to put captions on a video, and they are for different jobs.
A sidecar file (SRT or VTT) is a separate text file with timings. The player draws the captions over the video at watch time. Viewers can turn them off, resize them, change the language, and search engines and screen readers can read the text. YouTube, Vimeo and every video element on the web support them. This is the right form for anything longer than a minute and for anything where the viewer should have a choice.
Burned in (also called open captions) means the text is rendered into the pixels of the video itself. It cannot be turned off, resized or translated. It survives every upload, every re-share, every platform that ignores subtitle files, which is most of the short-form ones. For a 30-second vertical clip going to TikTok, Reels or Shorts, burned in is the only form that reliably arrives.
On this site the two are separate tools. The auto-captions tool burns captions into the video; the transcript tool writes the same recognition out as SRT or VTT. If you are unsure, produce both from the same recording: burn the captions into the file you post, and keep the SRT alongside it for anywhere that will take it.
How much text at once
Broadcast subtitling has settled these numbers over decades and they transfer directly. A line should be no more than about 37 to 42 characters. There should be no more than two lines on screen. Together that is a caption of at most about 80 characters, and most should be shorter. The reason is not aesthetic: the eye reads a caption in saccades, and a line longer than that needs more of them than a viewer can spend before the picture has moved on.
For vertical video the numbers are tighter. The frame is narrower, the text is larger, and the style that works is one short phrase at a time, three to six words, often with the word currently being spoken highlighted. That karaoke style is not decoration; it tells the viewer exactly where they are in the sentence without having to read the whole caption again every time it changes.
How long it stays
The guideline is a reading speed of about 15 to 20 characters per second for adults, with a minimum display time of around one second even for a single word, and a maximum of about six seconds before the caption should change whether or not the speaker has finished. A caption that flashes up for 400 milliseconds cannot be read. A caption that sits for eight seconds is read twice and then ignored, so the change to the next one is missed.
Speech recognition gets this partly wrong in a specific way. Models like Whisper, which the caption tools here run in your browser, time words well at their start and loosely at their end: the last word of a sentence is often stretched to cover the pause that follows it. If you trim a caption's end to where the word audibly stops, the caption vanishes while the speaker is still mid-word on the next one. The grouping on this site deliberately lets each line occupy its full span up to the next line's start, and distributes the word highlights within it, which is the behaviour that keeps every spoken word covered. Leave it alone unless a specific caption is obviously wrong.
Where it goes
On a vertical 9:16 video, the bottom fifth of the frame is covered by the platform's interface: the caption text, the username, the sound name, and on the right side the like and share buttons. The top has the status bar and sometimes a search bar. Text placed in either region is unreadable, and this is the single most common mistake on short video. The safe area is roughly the middle 60 percent of the height, with the text centred horizontally and kept away from the right edge. The caption tools here default to a bottom margin that keeps the text above that band on a 9:16 frame, and the margin is a control, so check it against the platform you are posting to.
On landscape video for YouTube or a web page, the bottom of the frame is the conventional place and it is fine, because the player's controls fade out.
Style
- A bold sans-serif, large enough that a line fills no more than about 80 percent of the frame width on a phone. Thin fonts disappear over footage.
- High contrast that does not depend on the background: white text with a dark outline or a drop shadow reads over anything. White on a bright sky, or black on a dark jacket, does not. A semi-opaque box behind the text is the broadcast solution and still works.
- Sentence case. All-caps is slower to read and is only worth it for one or two words of emphasis.
- No more than two styles in a video. Colour changes on keywords work; colour changes on every line become noise.
Fixing the words
Recognition errs in predictable places: proper names, product names, technical terms, numbers, and anything said over music. Read the transcript once before rendering. The transcript tool gives you the plain text to skim, and the caption editor on the auto-captions tool lets you retype a line and keeps the word timing spread across it. The things worth checking every time are names, figures, and the first and last lines, which are the ones a viewer is most likely to screenshot.
Punctuation is part of timing. A caption that ends mid-clause, with the rest on the next card, reads worse than one that breaks at a comma or full stop, even if the second is slightly longer. Where you have a choice of where to split, split where a reader would pause.
Subtitle files, briefly
SRT is the simplest and the most widely accepted: a number, a time range, the text, a blank line. VTT is the web's version, nearly the same with a header and support for positioning and styling. Both are plain text you can open and fix in any editor. ASS, which the burn tool also accepts, carries full styling, fonts, positions and effects, and is the format to use if you have designed the look somewhere else and want it rendered exactly. If a platform lets you upload a subtitle file, SRT is the one it will take.
The short version
Two lines at most, under 40 characters each, at 15 to 20 characters per second, for at least a second and at most six, in the middle of a vertical frame or the bottom of a landscape one, white with an outline, burned in for short-form and attached as SRT for everything else. Read the names. That is the whole discipline, and almost nobody does all of it.