How to translate video captions
Auto-translated captions inherit every error in the auto-transcript underneath them, and translated text runs longer than the English it replaced. Here is the order of operations that avoids both.
8 minute read
Translation is two problems stacked, not one
Almost every translated caption starts life as an automatic transcript, and that is where most of the damage happens. YouTube says as much on its own help page: automatic captions can misrepresent the spoken content because of mispronunciations, accents, dialects or background noise, and creators are told to review the output and edit anything transcribed wrongly. Translation does not fix any of that. It launders it. A misheard product name comes out the other side as a fluent, grammatical, entirely wrong sentence in Spanish, and nothing in the output flags it. A viewer in Madrid cannot tell a translation error from a transcription error, and neither can you once it is translated. So the first rule is boring and non-negotiable: correct the source transcript before you translate a single line.
How good machine translation actually is
The clearest recent read is the WMT24 general machine translation shared task, whose findings paper is titled The LLM Era Is Here but MT Is Not Solved Yet. The organisers covered eleven language pairs across several domains, collected output from eight large language models and four commercial online translation providers alongside the submitted systems, and had professional annotators mark errors span by span rather than score whole sentences. The conclusion is in the title. Modern systems are genuinely good and genuinely unfinished. It is also worth noticing where vendors and researchers diverge: near-human claims are usually measured on generic prose, while the shared task deliberately spreads across domains and grades how bad each error is. Your podcast is a domain, full of names and jargon.
Translated text does not fit the same box
Underneath the linguistic problem is a physical one. Translated text is usually longer than the English it replaced. The W3C's guidance on text size in translation reproduces IBM's expansion averages for English into European languages, and they are steeper than most people expect: source strings up to ten characters average 200 to 300% of their original length, 31 to 50 characters average 140 to 160%, and only past 70 characters does it settle around 130%. Set that against real subtitle limits. Netflix's English style guide allows 42 characters per line, a maximum of two lines per subtitle event, and a reading speed up to 20 characters per second for adult programming, 17 for children's. Same timing, more text, less time to read it.
Why burned-in translations behave differently from a subtitle track
These are not the same product. A subtitle track is data: one upload serves every language, the viewer picks, search engines and transcript panels can read it, and you can correct a bad line months later without touching the video. It is also optional, so plenty of viewers never turn it on, and the platform renders it in its own style, at its own size, wherever it decides. A burned-in translation is pixels. It always shows, and it looks exactly as you designed it, but you get one language per file, nothing machine-readable, and a single wrong word means a re-render. Expansion bites here in a way it does not bite a track: a line that fitted the frame in English can wrap to three lines in German and cover the speaker's mouth.
What YouTube does for free, and where that stops
YouTube gives away more of this than people realise, and the edges are worth knowing. Viewers can auto-translate any video that already carries a caption track, from the subtitles menu in the player, into a long list of languages. That costs you nothing and inherits every flaw in the track underneath it, which is one more argument for fixing the source. What YouTube took away was crowd translation: community contributions were retired in September 2020, with YouTube reporting that under 0.001% of channels had published community captions in the previous month. What it added is audio. YouTube's blog announced on 4 February 2026 that auto dubbing is open to all creators in 27 languages, with Expressive Speech in eight of them.
The order of operations that saves the most money
Work in this sequence and you will spend a fraction of what most people spend. Correct the source transcript first, concentrating on names, brands, numbers and anything with a unit attached, because those are the errors that survive translation looking confident. Translate from the corrected text, never from raw auto-captions. Then re-time rather than re-translate: since the target text is longer, the usual fix is splitting a cue or holding it a beat longer, not choosing different words. Have a native speaker check the first sixty seconds and the call to action rather than the whole file, since those two stretches carry most of the risk of reading as machine output. And choose languages from your own analytics, not from a list of large markets.
Where clips make this harder, and easier
Short vertical clips change the calculation, because there is no subtitle track to upload. TikTok and Instagram transcribe on the platform and publish no caption-file route for an ordinary post, so a translated clip means a separate render per language with the text burned in. That is only expensive if the transcript is locked. If your clipping tool keeps the transcript editable and re-renders from it, one corrected source becomes several language versions without re-cutting anything, which is how the caption step works in FrameOS. The honest caveat is that each translated clip is a new asset with its own description, thumbnail and comments to manage. That overhead, not the render, is the real cost.
FAQ
Is machine translation good enough for video subtitles?
For comprehension, usually yes. For anything load-bearing, not on its own. The WMT24 shared task, which evaluated eleven language pairs with professional annotators, published its findings under the title The LLM Era Is Here but MT Is Not Solved Yet. Treat machine output as a strong first draft, and have a native speaker check the opening and the call to action.
Should translated captions be burned in or uploaded as a file?
Upload a file for long-form, where one video can carry many languages, viewers choose, and you can fix errors after publishing. Burn them in for short vertical clips, where feeds autoplay muted and TikTok and Instagram offer no caption-file upload for a normal post. Burned-in means one render per language, so budget for that.
Does YouTube translate subtitles automatically?
Yes, on the viewer's side. If a video already has a caption track, viewers can pick auto-translate in the player's subtitles menu and read it in their own language. You do not have to do anything, and you do not control the result. The quality depends entirely on the caption track underneath, which is why correcting the source track matters more than adding languages.
Does YouTube's multi-language audio dub my video for me?
No, and the two features get confused constantly. Multi-language audio lets you upload dubbed tracks you produced elsewhere; it does not generate them. Auto dubbing is the separate feature that creates a track for you. If an automatic dub already exists on a video, YouTube requires you to delete it before uploading your own version in that language.
Sources
- Findings of the WMT24 General Machine Translation Shared Task: The LLM Era Is Here but MT Is Not Solved Yet · ACL Anthology
- Text size in translation · W3C
- English (USA) Timed Text Style Guide · Netflix Partner Help Center
- Unlocking a global audience with auto dubbing · YouTube Blog
- Use automatic captioning · YouTube Help
Related pages
Keep reading
More Captions & Subtitles →How to add subtitles in CapCut
CapCut transcribes a clip in seconds. The five minutes after that — corrections, line breaks, styling and placement — are what separate readable subtitles from a wall of text.
7 minute readWhat are burned-in captions?
Burned-in captions are subtitles baked into the video itself, so they always appear and can't be turned off. Here's when to use them versus closed captions.
5 minute readSubtitles vs captions: what's the difference?
Subtitles and captions are often used interchangeably, but they're not the same. Here's the actual difference — and which one you need for short-form video.
5 minute readTurn one long video into a week of posts
300 credits for 3 days · no card.
Start free