How to Translate Subtitles for Broadcast
What makes translating subtitles harder than translating text?
Subtitles are translated inside fixed constraints the source text never had: each event has a start time, an end time, a character-per-line cap, and a reading speed ceiling. Most European languages run 15 to 35 percent longer than English, so a faithful translation routinely will not fit the time it has.
That single fact explains almost everything that is unusual about audiovisual translation. A document translator can let a sentence grow. A subtitle translator cannot, because the timing belongs to the picture and the picture is not changing.
What has changed since this article was first written is the quality of the first draft. Machine translation and LLM-based translation now produce drafts good enough that the human effort has shifted almost entirely from translating to condensing. The workflow below reflects that.
The Four Constraints Every Translated Subtitle Has to Satisfy
Before any of the tooling matters, these are the rules the output is judged against. A translation that breaks any one of them will come back from QC regardless of how good the language is.
| Constraint | Typical broadcast value | What breaks it |
|---|---|---|
| Characters per line | 37–42 for Latin scripts | Compound nouns and untranslated proper names |
| Lines per event | 2 (occasionally 3 for CJK vertical) | Expansion forcing a third line |
| Reading speed | Set per language — see below | Same text, denser script |
| Timing | Inherited from the source, locked to picture | Nothing — you cannot move it to make room |
Our guides to subtitle reading speed and shot change rules cover the first three in depth for original-language work. This article is about what happens to them under translation.
Translation, Localization, and Adaptation
These three words get used interchangeably on purchase orders, and they are not the same job or the same price.
- Translation converts meaning from one language to another. The reference point is the source text.
- Localization adapts for the target culture — idioms, measurement units, cultural references, honorifics, register. The reference point is the target audience.
- Adaptation rewrites for effect where a direct rendering would fail: song lyrics, wordplay, jokes built on the source language, on-screen text that has to carry a pun.
Most broadcast subtitle work is localization with pockets of adaptation. Machine output is translation. That gap is the work, and it is why the post-editing pass is not optional on anything with a comedic or culturally specific register.
Text Expansion: The Constraint That Actually Bites
| Target language | Typical expansion from English | Practical consequence |
|---|---|---|
| German | 20–35% | Worst case. Compound nouns are long and cannot be hyphen-broken arbitrarily |
| Spanish | 15–30% | Consistent overflow on two-line events; condensing is routine |
| French | 15–25% | Elision shortens the audio but not the written form |
| Italian, Portuguese | 15–25% | Similar to French in practice |
| Japanese, Chinese, Korean | Contracts | Fewer characters, but much lower reading speed ceilings offset the gain |
These are planning figures, not guarantees — register and subject matter move them. The point is directional: if you budget a translation pass on the assumption that the text will be the same length, you have under-budgeted.
The practical mitigation is to condense at the translation stage rather than the QC stage. A translator who knows the character budget writes to it. A translator handed an unconstrained MT draft to "clean up" will tend to preserve the draft's structure and blow the limits.
Reading Speed Limits Change by Language
This is the detail most often missed when a workflow built for English is pointed at another language. The ceiling is not universal. Netflix's per-language timed text style guides, which much of the industry follows as a de facto reference, set markedly different values:
| Language | Adult programming | Children's programming |
|---|---|---|
| Traditional Chinese | 9 characters/second | 7 characters/second |
| Korean | 12 characters/second | 9 characters/second |
| Indonesian, Romanian, Slovak | 17 characters/second | 13 characters/second |
| Telugu | 22 characters/second | 18 characters/second |
Two consequences follow. First, a QC profile has to be per-language; a single 17 CPS rule applied to a Chinese delivery passes files that are unreadable. Second, CJK contraction is not the free win it appears to be — fewer characters against a 9 CPS ceiling can still overrun.
Always confirm the values in the delivery spec you were sent. Platform guides are revised, and a broadcaster's own house spec overrides any general reference.
Machine Translation, LLMs, and Human Translators
Three approaches are in active use, and the sensible answer is usually a combination rather than a choice.
| Approach | Strength | Weakness | Use it for |
|---|---|---|---|
| Neural MT (DeepL, Google, ModernMT) | Fast, cheap, predictable, glossary-enforceable | Segment-by-segment — limited view of surrounding dialogue | Volume first drafts, pivot generation |
| LLM translation | Better fluency and context handling across a scene | Can introduce confident errors not present in the source | Dialogue-heavy content where register matters |
| Human translator | Judgement on culture, register, adaptation, condensing | Cost and turnaround | Final pass on anything delivered; sole pass on premium drama |
The shift worth naming: neither machine approach reliably respects the constraints in the first table. They translate the words, not the box the words have to fit in. Reading speed violations, three-line events, and bad line breaks are the normal output of an unsupervised machine pass, which is why the post-edit is a subtitling task rather than a proofreading one.
Closed Caption Creator supports machine translation through nine providers — Amazon Translate, Anthropic, Azure AI Translator, DeepL, Gemini Pro, Google Translate, Lara AI, Modern MT and OpenAI — with the translated events landing on their own linked Event Group so the source stays intact. Our guide to choosing a machine translation engine for subtitling compares them on tag handling, glossaries and output length. See machine translation for what is included in each plan.
When to use a pivot language
If no translator exists for your pair — and for many pairs outside the major languages, none practically does — translate into a pivot language first, usually English, then out of the pivot into each target. It works, with one caveat: errors in the pivot propagate into every downstream language at once. Have the pivot reviewed before it becomes a source file, not after.
How to Translate a Subtitle File, Step by Step
- Lock the source. Translate from a finished, QC-passed original-language file. Translating against a moving source is the single most expensive mistake in the workflow — or use parallel translation deliberately, below.
- Get the target spec first. Characters per line, reading speed, line count, and file format for that language and that destination. Not the English one.
- Build or import the glossary. Names, places, invented terminology, recurring phrases. Enforce it in the engine rather than relying on translators to remember it.
- Generate the draft. Machine or LLM translation into a new linked Event Group, keeping the source timings and the source text side by side.
- Post-edit against the constraints. This is the real work: condense to fit reading speed, reflow line breaks for the target language's grammar, fix register, adapt what cannot be translated.
- QC with a target-language profile. Reading speed, line length, and character set checks configured for the target, not inherited from the source.
- Export per language in the format the destination asked for, with the language correctly declared in the file.
Glossaries and Key Names & Phrases
Consistency across episodes matters more than elegance in any one line. A character name rendered three different ways across a season is the complaint that reaches the broadcaster. That is what a Key Names and Phrases (KNP) document exists to prevent: an agreed list of how names, places, invented terms, and catchphrases are written in each target language.
The modernization here is worth acting on. KNP used to be a spreadsheet that translators were asked to consult. The major MT providers now expose glossary and terminology APIs — DeepL’s glossary feature among them — which apply the same list automatically at translation time, though which engines expose them, and how varies a great deal. The document becomes an enforced constraint rather than a request.
Three rules that keep a KNP useful:
- Own it centrally. One list per title, per language, not one per translator.
- Version it. When a rendering changes in episode six, you need to know which earlier episodes to revisit.
- Include what must not be translated. Brand names, on-screen signage, and sometimes song titles are as important as what must.
Parallel Translation of Work in Progress
Broadcast schedules often will not wait for a locked source. Parallel translation means starting the translation while the original language file is still being edited, and syncing changes down as they land.
In Closed Caption Creator, translated subtitles sit in an Event Group linked to the source group. When the original changes, Event Group options → Sync From Linked Group pulls new events and revised timings into the translation without discarding translated text that is still valid.
It compresses the schedule meaningfully. It also costs something: every sync is a small reconciliation, and translators are working against text that may change under them. Use it when the deadline requires it, and prefer a locked source when it does not.
Reviewing Multilingual Subtitles
Review is usually the stage with the least tooling and the most risk, because the reviewer frequently does not speak every language in the delivery.
What can be checked without speaking the language
More than people expect. Reading speed, characters per line, line count, event count against the source, timing identity with the source, character set and encoding, and whether any event is still in the source language. All of these are mechanical, and all of them are common failures. Our caption QC guide covers the mechanics; QC and review covers what the application flags automatically.
What needs a speaker
Register, honorifics, idiom, adaptation quality, and whether the condensing lost something that mattered. Budget for it on anything premium.
Getting the file in front of a reviewer
Not every reviewer will install a subtitle editor. Closed Caption Creator exports translations to Microsoft Word, Excel, and plain text and re-imports the edited file, which keeps the review in a tool the reviewer already has. Excel is the practical choice for multi-language review — one column per language against a shared timecode column shows inconsistencies that are invisible when each language is reviewed alone.
What You Actually Deliver
| Destination | Typical format | Watch for |
|---|---|---|
| European broadcast | EBU-STL or EBU-TT | Code page limits on STL will drop characters outside the selected set |
| Streaming platforms | IMSC / TTML | Language declared in the file; one file per language |
| Web and OTT players | WebVTT | Segment alignment in HLS; label and language tags in the manifest |
| US broadcast | SCC or MCC for the captions, sidecar for foreign-language subtitles | SCC character set will not carry most non-English diacritics |
| Review and sign-off | Word, Excel, plain text | Not a delivery format — re-import before export |
The encoding row is the one that catches people. A translation that is correct in the editor can lose its accented characters on export to a legacy format, and it will not be obvious until someone reads the delivered file. Export early, open what you exported, and check. Our streaming delivery guide goes through the platform-by-platform requirements, and format support lists what reads and writes.
Frequently Asked Questions
What makes translating subtitles harder than translating text?
Subtitles are translated inside fixed constraints the source text never had: each event has a start time, an end time, a character-per-line cap, and a reading speed ceiling. Most European languages run 15 to 35 percent longer than English, so a faithful translation routinely will not fit the time it has.
Can you machine translate a subtitle file?
Yes, and it is standard practice as a first pass. Machine translation and LLM translation both produce usable draft subtitles in seconds. What neither does reliably is respect reading speed, line length, and line-break logic in the target language, so a human pass to condense and reflow the text is still required for broadcast delivery.
How much longer does translated subtitle text get?
Translating from English, expect roughly 15 to 25 percent expansion into French, 15 to 30 percent into Spanish, and 20 to 35 percent into German. One engine, Amazon Translate, has a setting built to shorten the output for exactly this reason. German is the hardest case because compound nouns create long strings that cannot be broken across lines cleanly. CJK languages generally contract instead.
What is a pivot language in subtitle translation?
A pivot language is an intermediate language used when no translator is available for a direct language pair. The source is translated into the pivot, usually English, and then from the pivot into each target language. It makes rare pairs workable but compounds error, so the pivot file should be reviewed before it is used as a source.
What is a KNP document in subtitle localization?
A Key Names and Phrases document lists how character names, place names, invented terms, and recurring catchphrases must be rendered in each target language. It keeps episodic content consistent across episodes, seasons, and translators. Modern machine translation glossary APIs can enforce the same list automatically at translation time.
Do translated subtitles use the same reading speed limit as English?
No. Reading speed ceilings are set per language, because character density differs. Netflix per-language guides allow up to 17 characters per second for adult Latin-script subtitles but only 9 for Traditional Chinese and 12 for Korean. The English guide permits 20 for SDH and English subtitles.
What file format should translated subtitles be delivered in?
It depends on the destination. European broadcasters commonly take EBU-STL or EBU-TT, streaming platforms generally want IMSC or TTML, and web delivery uses WebVTT. One subtitle file per language is the normal unit of delivery, and the language must be declared correctly in the file or its filename.
Running a multi-language delivery? See the translation workspace or talk to our team.