Machine-translated subtitles: what to check before you publish

29 September 2026 · 6 min read

A machine-translated subtitle file is a good first draft, and on informative speech it is often close to final. The errors that remain are not random. They cluster in a few places, which means a ten-minute check catches most of them without re-reading the whole file.

1. Lines that grew

Translation almost always makes text longer. The W3C's internationalization guidance, citing IBM's figures, puts the expansion from English at around 30% for longer strings, and far more for short ones: a phrase under ten characters can double or triple. Subtitles are exactly the short strings that grow most.

A cue that was comfortable to read in English can become too long to read in the time it is on screen. Run the translated file through a reading-speed check and look at the cues that fail. The fix is to shorten the wording, not the timing: the timing is tied to the speech.

2. Names and terms

Models translate words that should be left alone: a person called Rose, a product called Notion, a technique with an English name that the target audience uses in English anyway. Search the translation for every proper noun in the original and check each one. This is the single most common error and the most embarrassing to publish.

3. Idioms, jokes and wordplay

A literal translation of an idiom is usually grammatical and wrong. Jokes that depend on a double meaning cannot survive at all. Find the lines where the speaker is being playful and rewrite them for the target audience, or translate the meaning plainly rather than the words.

4. Formal and informal register

Korean, Japanese, German and French all mark formality in ways English does not. A model picks one, and it may not be the one your channel uses. Read the first few cues aloud in your head: if a casual vlog sounds like a press release, fix the register once and apply it through the file.

5. Line breaks in Chinese and Japanese

Chinese and Japanese are written without spaces, so a line break can land in the middle of a word. Each character is also roughly twice the width of a Latin one, so a line holds fewer of them. If the translation is going into a CJK language, check where the long cues wrap.

A ten-minute routine

1. Run a reading-speed check and fix every cue that fails.
2. Search for each name and term from the original.
3. Read the first thirty seconds for register.
4. Jump to the moments you know are playful and rewrite them.
5. Play the video with the subtitles on and watch for anything that feels off.