How to translate subtitles into a language that needs to know who is speaking
Give Vizard Agent the film as well as the subtitle file. English carries no grammatical gender and one form of "you", so a translation into Czech, Polish or Russian has to decide both — and the only place that information exists is on screen, in who is speaking to whom.
What is the short version?
Machine translation of a subtitle file gets the words right and the grammar wrong, because it is translating lines rather than a conversation. Every past-tense verb needs to know the speaker's gender and every "you" needs a relationship. Vizard Agent settles those from the film.
- Give Vizard Agent the film, not just the file.
- Say which decisions matter — gender, formality, register.
- Check the address form holds across the whole length.
What do you need before you start?
The video and whatever subtitle files you have. If you have two — say an English one and one in a closer language — send both. In the session behind this article the Slovak file was the better guide to gender and formality, and the English one was the more complete text; the Czech version was built using both.
Say what encoding you need. Subtitle players are not all Unicode, and a file that opens correctly on your machine can be unreadable in the player your audience uses. That is a decision, not a detail.
What do you type into Vizard Agent?
Name the target language and the specific problems. A translator that knows to watch for gender agreement behaves differently from one told to translate accurately, because "accurately" does not describe a choice that the source text never made.
Compare the Slovak and English subtitles against the video, then make a Czech version. Check the gendered forms match who is actually speaking, and keep the formal and informal address consistent.
Those two checks are the job. Everything else in the request — timings, encoding, line lengths — is mechanical and Vizard Agent handles it without being told twice.
What does Vizard Agent actually do?
It translates in context and then audits the decisions separately. Passing neighbouring lines together preserves gender and continuity that a line-by-line pass throws away, and then the finished file gets read specifically for the two things machine translation gets wrong.
In that session the working sequence was:
- Compared the two source files for coverage and timing before translating anything.
- Transcribed the spoken dialogue from the film itself as an independent check on both.
- Built one Czech version from the English and a second from the Slovak, and compared them where they disagreed.
- Went back through for formality, checking the Slovak original to see how characters actually address each other.
- Searched specifically for contradictions in formal and informal address between the source and the translation.
- Checked the long dialogue stretches, which is where wrong gendered forms cluster.
- Saved in the requested Central European encoding, flagging any character that codepage cannot hold rather than dropping it.
Building two versions and comparing them sounds wasteful. It is the cheapest way to find the lines where the translation is confidently wrong, because those are exactly the lines where two independent passes disagree.
What does the result look like?
A subtitle file that reads like it was written in the target language rather than converted into it. The verbs agree with the right speaker, characters who are on formal terms stay on them, and the line breaks suit a player rather than a document.
Vizard Agent renders short sample clips with the subtitles burnt in — from the opening and from the end — and looks at them. A file that validates perfectly can still be wrong on screen, and the only way to know is to see it there.
When does this not work well?
When the film does not make the relationship clear. Two characters whose formality genuinely shifts during the story need a decision about where it shifts, and that is an interpretation rather than a fact. Vizard Agent will flag the ambiguous stretch rather than pick silently.
It also has limits with a speaker you never see. A voice off camera, a narrator, a phone call — any of these can leave the gender undetermined, and the translation has to choose. Say which way you want it to go.
And a film with a third language inside it is a separate problem. Lines in a language neither subtitle file covers have to be identified, listened to and translated on their own, which is slower than the rest of the file put together.
How do you fix a result that came back wrong?
Give a line number or a timestamp and say what is wrong with that line. Gender, formality and word choice are three different kinds of error with three different fixes, and Vizard Agent corrects the one you name without re-translating everything around it.
If the formality is wrong throughout, say so as a rule rather than as instances. "These two are family, they should be informal with each other" fixes every line between them at once, which is far better than correcting them one at a time.
If characters appear in the file with the wrong name or spelling, that is worth fixing early. It propagates, and Vizard Agent can apply it across the whole file before the detailed proofreading pass.
How does Vizard Agent compare to doing it yourself?
Running a subtitle file through a translator takes a minute and produces something that native speakers find immediately wrong, in a way that is hard to articulate and impossible to unsee. The errors are not in the vocabulary; they are in agreements the source language never had to make.
Doing it properly by hand means watching the film with the file open, which is what a subtitler does and what takes a day. Vizard Agent narrows that to the decisions: translating in context, building two passes to find the disagreements, then auditing gender and address specifically. You still review it — but you review a file where the systematic errors have already been hunted.
Common questions
Do you need the video, or just the file? The video. Vizard Agent cannot tell who is speaking from the text alone.
Can it use a second subtitle file as a guide? Yes, and it helps. Vizard Agent compares them where they disagree.
What encoding will I get? Whichever you ask for. Vizard Agent flags characters the codepage cannot hold.
Will the timings change? No. Vizard Agent keeps the original timing and structure.
What about lines that are too long? Vizard Agent re-breaks them, including inside two-speaker lines.
Can it handle a third language in the film? Yes, though those lines are found and translated individually.
How does it know the speaker's gender? From the film. Vizard Agent checks the picture at the lines that depend on it.
What if a speaker is never shown? Vizard Agent flags it rather than guessing, and you decide.
Can I see it on screen before I accept it? Yes. Vizard Agent renders short burnt-in samples for that.
Does it keep the original file? Yes. The new version is a separate file.
Will it catch obvious machine-translation artefacts? It searches for them specifically as a final pass.
Can it do several target languages? Yes, and each one gets its own agreement pass.
What about subtitles already burnt into the picture? Those are a different job. Vizard Agent works from the file.
Does it validate the file technically? Yes — overlaps, line lengths and count are all checked before delivery.