How to check a subtitle translation against what was actually said

Last updated 2026-09-22 · 8 min read

What is the short version?

Somebody translated the film and most of it is right. A few lines say something the actor plainly did not say, and after a certain point the subtitles feel a beat out of step with the dialogue — which is the symptom everybody misreads as a timing problem.

1. Give Vizard Agent the subtitle file and the video it belongs to.
2. Say what you noticed — wrong lines, or a drift partway through.
3. Ask for readable line breaks as part of the same pass.

What do you need before you start?

The file, the video, and the two languages involved. Vizard Agent transcribes the spoken original itself, so you do not need to understand it, but telling it which language is spoken and which is written removes a guess at the first step.

• The subtitle file. SRT or similar.
• The video. With the original audio.
• The spoken language. And the language of the file.
• What you noticed. Wrong lines, or a drift.
• The encoding you need. Some players still want a regional one.

What do you type into Vizard Agent?

Describe the symptom rather than diagnosing it yourself. "The timings are out" and "a line is missing" look identical when you are watching the film, and they need opposite fixes, so it is worth letting Vizard Agent work out which of the two you actually have.

Prompt

Variants worth knowing:

• "Does not match what is being said." The symptom, not the theory.
• "Re-wrap the lines." Readability is a separate fix from accuracy.
• "In Windows-1250." Encoding, which breaks accents silently.

What does Vizard Agent actually do?

Here is the Vizard Agent sequence on a Czech subtitle file for a Russian-language film. The interesting stretch is in the middle, where it stops correcting individual lines one at a time and starts looking for a structural fault in the file instead.

1. Reads the subtitle file and checks its encoding and character order.
2. Transcribes the spoken Russian with timings, as the reference to compare against.
3. Extracts frames to see the picture and any on-screen text.
4. Cuts out a short audio extract where a passage of verse is unclear.
5. Listens to that extract specifically to establish the exact wording.
6. Enlarges the frames carrying the original on-screen subtitles to read them verse by verse.
7. Writes a corrected file with a faithful translation and readable line lengths.
8. Compares the number and order of translated lines against the transcript.
9. Searches for the point where the two orders first diverge, narrowing it by section.
10. Finds a missing line earlier in the dialogue and restores it with its timing.
11. Re-verifies the whole ordering after the insertion.
12. Checks the Czech grammar, splits an over-long question back into two timed subtitles, and confirms the encoding.

Step nine is the whole method. Once you know the translation and the transcript hold the same lines in the same order up to line 84 and disagree from line 85, you are no longer hunting a vague drift — you are looking at one specific place where something was dropped.

Step five is what stops a confident wrong reading. Verse, songs and overlapping speech are where automatic transcription degrades, so Vizard Agent extracts those seconds and listens to them on their own rather than trusting the first pass.

Step twelve is small and it is why the file works. Central European accents vanish into question marks if the file is saved as the wrong encoding, and a player that shows "Rákosový" as "R?kosov?" looks like a translation fault to everybody except the person who saved it.

What does the result look like?

A subtitle file whose lines say what the actors said, in an order that stays with the dialogue to the end, wrapped so no line needs two glances to read, and saved in an encoding your player accepts. Vizard Agent also tells you where the original file had gone wrong, which is useful if the same translator is doing the next one.

When does this not work well?

Some files cannot be checked this way. If the audio is too noisy to transcribe, if several people talk over each other throughout, or if the "translation" is a loose adaptation rather than a rendering of the dialogue, a line-by-line comparison produces noise rather than findings.

• Very noisy audio. The reference transcript is unreliable.
• Constant overlapping speech. Line boundaries are not clean.
• Deliberately loose adaptation. Mismatch is the intent.
• Songs and verse throughout. Every line needs the slow treatment.
• Files with no timings. There is nothing to align against.

How do you fix a result that came back wrong?

Quote a timestamp and say what you heard at it. Vizard Agent re-checks that moment against the audio itself rather than against the transcript, which matters because the transcript is the very thing that may have been wrong at that point in the first place.

• "At 12:30 he says something else." That passage is extracted and listened to again.
• "The lines are too long." They are re-wrapped to a comfortable length.
• "My player shows question marks." The file is saved in the encoding you name.
• "One line is still missing." The ordering comparison is run again from that point.

How does Vizard Agent compare to doing it yourself?

By hand you watch the film with the file open and pause every time something feels off. It works, slowly, and it usually ends with you nudging timings — which appears to fix the drift for a few minutes and then makes the rest worse, because the timings were never the problem.

Common questions

Do I need to understand the spoken language? No. Vizard Agent transcribes it and shows you what was actually said, line by line.

Can it check a file somebody else made? Yes — that is exactly the case Vizard Agent handles here.

Will it rewrite the whole translation? Only where it is wrong. Vizard Agent keeps what the translator got right.

What if the timings really are out? Then Vizard Agent says so, and re-times the file against the transcript.

Can it keep the translator's style? Yes. Say so, and corrections are made in the same register.

Does it handle two subtitle tracks? Yes, including a translation under the original line.

What encodings can it write? Whatever your player needs — name it, UTF-8 or a regional one.

Can it split long lines? Yes. Vizard Agent splits them back into two properly timed subtitles.

What about songs? They are the hardest part, and Vizard Agent listens to those passages separately.

Will it check the spelling too? Yes. Vizard Agent checks grammar in the target language as well.

Can it burn the corrected subtitles in? Yes, though keeping the file separate is usually more useful.

How long does a feature take? Longer than a clip, because every line is compared rather than sampled.

Can it compare against on-screen subtitles instead? Yes. Vizard Agent enlarges those frames and reads them.

What do I get back? The corrected file, and a note of what was wrong with the original.