Managing Subtitles in Eight Languages Without Losing Track

The first translated subtitle file is easy. The eighth is where the work changes character: you are no longer producing subtitles, you are maintaining a set of eight things that have to stay identical in structure while being different in content, across every future correction to the video.

What fails at that scale is rarely the translation. It is a file with an ambiguous name that nobody can match to a language, a correction applied to six versions out of eight, a cue that reads fine in the source and runs at an impossible speed in German, and a re-cut of the video that silently invalidates everything. This guide covers the conventions and the ordering that prevent those, and the one decision that costs the most if you get it wrong.

Lock the Source First

Every downstream file is a copy of the source's structure. The cue boundaries, the count, the timings and the segmentation all propagate, which means a change to the source after distribution is a change you now make in every language, by hand, correctly, in a script you may not read.

So the source has to be finished before anything leaves. Finished means the wording is final, the segmentation is set, the timings are checked against reading speed, and someone has proofread the names and figures. That is a higher bar than "good enough to publish in one language", and clearing it is the single highest-value habit in multi-language work.

Treat the locked source as a released artefact. Give it a version, store it where it will not be casually edited, and if it genuinely has to change, bump the version and re-issue rather than editing in place. The correction pass that precedes locking — names, terms, numbers, punctuation — is the same one described in the order to correct an automatic transcript.

Name Files by Tag, Not by Word

Filenames are the only metadata that survives every tool, every hand-off and every zip file, so they have to carry the identifying information themselves.

Use standard language tags. es is Spanish, es-419 is Latin American Spanish, pt-BR is Brazilian Portuguese, zh-Hans is Simplified Chinese. These are unambiguous, they sort predictably, and they distinguish regional variants that a plain language name cannot — a distinction that matters, because a Spanish file for Spain and one for Latin America are genuinely different deliverables.

Match the video's base name and put the tag immediately before the extension:

episode-04.en.srt
episode-04.de.srt
episode-04.es-419.srt
episode-04.pt-BR.srt
episode-04.zh-Hans.srt

Two properties make this worth the discipline. A folder of forty files sorts into readable groups, and many players read exactly this pattern to build their subtitle menu without any further configuration.

Text Expansion Changes Your Timings

Translation changes length, and length was one of the inputs to your timing decisions. A cue given three seconds because the English line was 40 characters is still three seconds when the German line is 55, and the reading speed has changed underneath it.

The direction varies by language pair. Several European languages run longer than English for the same content, with German and Finnish the usual examples, and compound words that cannot be broken across lines. Languages written in Chinese, Japanese and Korean characters carry more meaning per character, so the character count drops while the reading time per character rises — which means a raw characters-per-second figure is not comparable across scripts and needs its own target per language.

The practical consequences are concrete. Line-length limits have to be re-checked per language rather than inherited. Cues that were comfortable may now exceed the reading budget, so either the translation is condensed or the cue is extended into an available gap. And a language with long unbreakable words will occasionally need a different line break than the source, which is a legitimate deviation rather than an inconsistency. The measurements that govern all of this are the same ones described in the timing rules for readable subtitles.

Brief the translator on the constraint rather than fixing it afterwards. A translator told the line limit and the cue duration produces text that fits; one given a flowing transcript produces text that reads beautifully and has to be cut down by someone who does not speak the language.

The QA Pass Order

Run technical checks before linguistic ones. Technical faults are cheap to find and they invalidate the linguistic review, so discovering a broken file after a native speaker has reviewed it wastes the expensive resource.

Track Status Somewhere You Will Actually Look

Eight languages times several states is more than anyone holds in their head, and the failure mode is not chaos — it is quietly shipping seven of eight.

Keep one table, with one row per language, and columns for the states each file passes through: translated, technically checked, linguistically reviewed, delivered. Keep it next to the files rather than in a separate tool, so updating it is not a context switch. The specific tool matters far less than whether it is close enough to the work that it stays current.

Record two things beyond status. Which source version each translation was made from, because that is what tells you whether a correction has propagated. And who reviewed each language, because the next correction should go back to the same reviewer.

Re-Cutting After Translation

This is the expensive one. If the video is re-edited after translation — a shot trimmed, a section removed, a pause tightened — every timing in every language moves, and there is no automatic repair. Shifting cues wholesale works only for a change at the very end; anything in the middle re-times everything after it.

So the sequencing rule is blunt: the picture locks before the source subtitle locks, and the source locks before translation starts. Where a re-cut genuinely cannot be avoided, it is cheaper to re-run the timing pass against the new cut and re-fit the existing translated text than to re-translate, but it is cheaper still to have finished the edit first.

The same ordering applies with more force if the project will also be dubbed, since audio has less tolerance for re-timing than text does — the stages and their dependencies are set out in how AI dubbing works step by step. And if you have not yet committed to subtitles for every market, the choice between reading and hearing is worth making before this whole pipeline starts, which is the subject of choosing dubbing or subtitles per market.

Next Steps

Open the folder holding your last multi-language delivery and try to answer one question from the filenames alone: which of these is the Latin American Spanish file, and which source version was it made from? If you cannot, the naming is carrying no information and the next correction will go to the wrong file.

Rename to language tags, write down the source version each file came from, and lock your next source properly before it goes out. Those three habits remove most of the failure modes above. The formats those files can be delivered in, and what each one supports, are covered in SRT, VTT and ASS compared, with the rest of the caption topics under captions and subtitles.