Guide· 5 min read

Automatic captions with AI: how they work, what goes wrong

Automatic captions, often called auto captions, are made in two steps. First, an AI writes down what you said and notes the exact time of each phrase or word. Then that text gets a design and is placed over the video. This guide explains the process in plain words, the most common mistakes, and how to ask for captions when an AI agent edits your video with EditorBrain.

EditorBrain · Updated

Hand-drawn diagram in three steps: a voice, a sheet of text with a clock next to each line, and a phone with the caption "hello everyone" over the video. Below, the word "Cloud" crossed out and corrected to "Claude".
From your voice to the caption, with the most common mistake. Illustration by EditorBrain, made with AI.

How they are made

An AI writes down what you said with its times, and then that text gets a design.

What goes wrong

Names, mixed languages, timing and text over the face.

What you ask for

The format, the treatment and the names that must be spelled right.

How automatic captions work

Two hand-drawn caption bars. The top one, by phrase, has a single clock for the whole bar. The bottom one, by word, has a clock under each word and one word highlighted.
Word-by-word highlighting needs a time for every word.

First, a speech recognition model listens to the audio and writes down what you said. Next to the text it stores when each phrase or each word starts and ends. That is a transcript with timestamps.

Then the text is split into short blocks and given a look: font, size, color and position. Word-by-word highlighting, often called karaoke, needs a time for every word and not only for every phrase.

Burned into the video or kept in a separate file

Burned-in captions are part of the picture. They look the same on every platform and the viewer cannot turn them off.

Hand-drawn comparison. On the left, a phone with the caption inside the picture: burned in. On the right, the same phone with no text and, beside it, an SRT sheet with lines and clocks: separate file.
Inside the picture or in an SRT file.

The other option is to export a file, such as SRT, that the platform displays on top of the video. An SRT file holds the text and the times, but not the font, the color or the position.

What usually goes wrong

A phone with a person on screen. The caption bar sits right over their mouth and a note reads "covers the mouth".

Names. Recognition writes whatever sounds closest, so "Claude" can come out as "Cloud". It happens with brands, surnames and technical terms.

Mixed languages. If you speak Spanish and drop in English terms, those words can come out misspelled.

Timing. If the video is cut after it was transcribed, the original times no longer match what you hear. It also shows when words are spread evenly instead of measured.

Text over the face. A caption can sit well in the first shot and cover your mouth once the framing changes.

How captions are made when an agent edits with EditorBrain

An open notebook with the EditorBrain mark and the title "glossary". On the page, the name "Claude Code" written with a check mark.

EditorBrain does not edit the video. It gives your agent, Claude Code or Codex, editing instructions and your preferences, and the agent produces the captions with the tools on your computer. In your brand you save whether you want captions, the treatment (clean, highlighted keywords, word by word or with a background), the position and the language.

To transcribe, the agent uses the route you choose: Tempo, the transcription built into EditorBrain, which uses credits, or a transcription tool you already have. The agent keeps the measured times and maps them through the cuts. If only phrase times exist, it makes phrase captions instead of inventing a time for each word.

For names there is your brand glossary, which Tempo uses to recognize them better, although it does not ensure the spelling. When the agent corrects a word by hand, it offers to add it to the glossary and saves it only if you agree.

How you review and correct

Before delivery, the agent checks the captions at final size. It makes sure they do not cover faces or logos and that they stay inside the safe areas, the parts of the screen the app does not cover with buttons. It repeats the check after every change of framing. That check does not replace yours.

You watch a preview on the review page. There you read the transcript and leave a comment at an exact moment, for example: "At 00:12 the text covers my face. Move it up and keep the cuts and the voice as they are". If you have the EditorBrain program installed on your computer, the agent can also export the captions as a separate file, in SRT or ASS format.

How to ask for AI captions: a request you can copy

You do not need technical terms. Say the format, the length, the treatment and the names that must be spelled right. Say which part cannot be lost, too. Below the request is what happens after you send it.

Ask your agent like this
Use EditorBrain to edit this video as a vertical Reel of about 40 seconds.
Add English captions, word by word, in my brand font and colors.
Transcribe with real word timestamps and check the names: "Claude Code" and "EditorBrain".
Keep my whole explanation and make sure the text never covers my face.
Show me a preview before exporting.
  1. The agent tells you how it will transcribe. If that route uses credits, it explains the cost and waits for your go-ahead.

  2. It transcribes your voice with a measured time for each word and checks how the names were spelled.

  3. It cuts the video and maps the caption times through those cuts.

  4. It places the text in your font and colors and checks that it does not cover faces or logos.

  5. It shows you a preview on the review page. You comment on the exact moment you want changed.

If captions are all you need

If your video is already edited and only the text is missing, a browser editor handles that with nothing to install. Kapwing generates word-by-word subtitles and lets you export them burned in or as an SRT, VTT or TXT file. Tella turns the transcript of its recordings into captions you can edit, burn in or download as SRT. An agent with EditorBrain makes more sense when captions are one part of a full edit in your brand.

Common questions

What do I do if a caption misspells a name?

Correct it in the review and ask to keep the timing of that word. With EditorBrain, the agent offers to add the right spelling to your brand glossary for later transcriptions with Tempo. The glossary helps recognition, but it is still worth checking names every time.

Which is better, burned-in captions or an SRT file?

It depends on where you publish. Burned-in captions keep your font, your colors and the position on every platform. An SRT carries only text and times, and the platform decides how it looks. You can have both: the video with burned-in captions and the separate file.

Do captions use EditorBrain credits?

Only if you transcribe with Tempo, the built-in transcription, which is quoted in credits. You can also use your own transcription tool or a transcript you already have, even on the free plan. Outside services may have their own cost, and the Claude or Codex subscription is paid separately.

Can I ask for word-by-word captions?

Yes, when the transcript includes a measured time for each word. If it only has phrase times, the agent can make phrase captions or measure the time of each word with a tool you have authorized. What it should not do is spread the words evenly.

Sources

Facts about other products come from their own official pages, read on November 16, 2026. Prices and limits change; check the vendor's page before deciding.

  1. Kapwing: caption generator
  2. Tella: features

Keep reading.

All guides