How to Add Auto Captions to TikTok Videos

Automatic captions have gone from a novelty to a default in about three years. That is a problem, because the default result is a verbatim transcript in a default style, and a verbatim transcript is not what captions are for.

This guide is about the step after generation: what to fix, in what order, and how to set up a caption style once so that every future video needs five minutes of work instead of thirty.

What auto captions actually produce

The model outputs a transcript with word-level timings. That is all. Everything that makes captions watchable is decided afterwards, by you or by a preset:

• which words exist (filler, false starts, repeated words)
• how the text is broken into lines
• how long each line is on screen
• where it sits in the frame
• how it is emphasised
• whether it is coloured, animated, or boxed

A video with a transcript that is 98% accurate and a style that is unreadable is worse than a video with 90% accuracy and a clean style. Accuracy is the easy part now; presentation is where captions succeed or fail.

The correction pass, in priority order

Do these in this order, because later steps depend on earlier ones.

1. Names and nouns. The most damaging errors, because a wrong name is instantly visible to the one person who cares and unfixable in comments. Check every proper noun.
2. Numbers. "Fifteen" versus "fifty" changes the meaning, and both are plausible predictions.
3. Domain vocabulary. Every niche has words the model has never seen in context. Fix them once and add them to a custom vocabulary if your tool supports it.
4. Filler and false starts. Remove "um", "uh", repeated words, and abandoned sentence openings. This is the step that turns a transcript into captions, and it is also the step most creators skip.
5. Sentence breaks. Break long spoken sentences into readable lines. A caption line is not a sentence; it is a unit the eye can take in during the time it is on screen.
6. Emphasis. Mark the words that carry the argument, and leave everything else plain.

If you only have time for two of those, do filler removal and sentence breaks. They produce the largest visible improvement per minute spent.

Setting the style once

The reason caption work feels endless is that most creators re-decide the style for every video. Decide it once, in this order, and save it.

Position. Lower third, clear of the app's interface. The exact margins are in vertical video safe zones, and they differ between TikTok, Reels and Shorts in ways that matter.

Line length. Around 32 to 38 characters. Two lines maximum. Longer lines force the eye to jump back and lose its place.

Weight and contrast. Bold enough to read over movement, with a shadow or stroke rather than relying on pure white. White text on a bright frame is unreadable in daylight, which is where most feed browsing happens.

Animation. One movement per line. A pop on each new line is enough. Animating every word makes reading impossible at speed.

Case. All caps for two or three words; sentence case for anything longer.

Colour. One accent colour for emphasis, matching your brand. More than one accent reads as noise.

Save that as a preset and never open the styling panel again unless you are deliberately changing direction. Style decisions made per video are how a channel ends up looking like five different channels.

Reading speed is the hidden constraint

The caption has to be readable in the time it is on screen. A useful working rule: the average viewer reads about 200 to 250 words per minute for comfortable comprehension, which is roughly 3 to 4 words per second at a glance. If your speaker is doing 180 words a minute and you are showing six words at a time, each chunk needs about 1.2 seconds on screen to be comfortable.

Where this breaks:

• Fast speakers. The captions outrun the reading. Fix by shortening lines further, not by speeding up the display.
• Long pauses. Captions that linger feel stuck. Cut them on the pause.
• Numbers and URLs. These need double the hold time; nobody parses a URL at 4 words per second.

If you want the deeper version of this, including how caption timing interacts with retention, see caption readability.

The transcript is also your script

The least-used benefit of auto captions: the corrected transcript is a clean, searchable record of what you said. Three uses that pay for the correction time.

• Repurposing. The corrected transcript is a shot list. Scan it for the turns and you have your clips: repurposing long videos.
• Descriptions and titles. The strongest sentence in the transcript is usually a better title than anything you would write cold.
• Search. Your own archive becomes queryable by phrase, which matters once you have fifty videos.

Common failures and their causes

| Symptom | Cause | Fix | | --- | --- | --- | | Captions lag behind the audio | Whole-file timing drift | Re-time from a clear marker, or re-run generation on the audio alone | | Captions flash too fast to read | Too many words per chunk | Shorten lines, hold longer | | Captions cover the subject's face | Fixed position, moving subject | Move the block per shot, or use a style that sits lower | | Text disappears behind the interface | Ignored safe zones | Raise the baseline | | Shifted captions after a trim | Captions not linked to the cut | Re-link or regenerate after the final cut | | Names consistently wrong | No custom vocabulary | Add the terms once | | Different look on every video | No saved preset | Build one preset and reuse it |

Captions are not subtitles

One distinction worth keeping straight, because it changes the writing: subtitles exist to make audio accessible, so they are verbatim and complete. Captions in short-form also exist to hold attention, so they are edited: filler removed, sentences tightened, emphasis added. Accessibility is still the floor, which is why removing filler is safe (the meaning is intact) and removing information is not.

If you are captioning for accessibility compliance rather than for the feed, keep the verbatim transcript available as the description or as a separate file. The two jobs have different outputs from the same source.

The five-minute routine

Once the style is saved, the whole job looks like this:

1. Generate captions on the final cut, not the rough cut.
2. Fix names, numbers, and domain words.
3. Strip filler and false starts.
4. Break long lines.
5. Emphasis on the load-bearing words only.
6. Watch once, on your phone, sound off, in daylight.
7. Export.

Seven steps, five minutes, and it is the difference between captions that help and captions that make the video feel automated. For the export side, export settings for vertical video covers bitrate and the compression artefacts that ruin burned-in text.