How to correct subtitles and have the translation follow your corrections

Last updated 2026-09-10 · 8 min read

What is the short version?

Automatic subtitles get names, dialect words and half-heard lines wrong, and only you can fix them. If the translation is generated from the audio it will get the same things wrong again, in a language you may not be able to check.

1. Go to Vizard Agent with the video and your corrections.
2. Send them as timestamps with the corrected line.
3. Ask for the translation after the source is right.

What do you need before you start?

The video and your list of corrections. That list is easier to write than it looks — a time, then what the line should actually say, and a marker on the ones that Vizard Agent should remove entirely rather than rewrite into something else.

• The video with its current subtitles.
• The corrections. Time, then the corrected line.
• The deletions. Marked as such.
• Where the subtitles should sit. Bottom, usually.
• Which language to translate into, and when.

What do you type into Vizard Agent?

Write it out as a list and mark the deletions differently from the edits. Mixing "change this line" and "remove this line" together in the same prose is exactly how a line you wanted gone comes back from Vizard Agent reworded instead.

Prompt

Variants worth knowing:

• "(remove this line)." Distinguishes a deletion from an edit.
• "Once that is right, translate." Sets the order.
• "The name is spelled..." Names are the usual casualty.

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real recording whose subtitles needed both a list of corrections and a translation afterwards. The order here is the entire point: nothing gets translated at all until the source subtitles are the text you actually approved.

1. Reads the existing subtitle track and its timings.
2. Applies each correction at the timestamp you gave.
3. Removes the lines you marked rather than rewording them.
4. Fixes the spellings, including names.
5. Moves the subtitles to the position you asked for.
6. Renders and checks the corrected source subtitles.
7. Translates from the corrected text, not from the audio.
8. Applies your corrections to the translation where you send more.
9. Checks the translated lines fit the frame and the timings.

Step seven is the one that saves the second round. A translation generated from the audio hears the same unclear word you just fixed and gets it wrong the same way, so your correction has to be made twice — once in each language, and the second time in a language you may not read.

Step nine is the check the first language never needs. A translated line is rarely the same length as the line it replaces, so the fit against the frame and against the original timing has to be looked at again even though the timings themselves did not move.

Step three is worth being explicit about. A line marked for deletion and a line left blank are different things in a subtitle file, and a request that only says "this is wrong" tends to produce a rewrite when what you wanted was silence.

What does the result look like?

Subtitles that say what was actually said, with the names spelled properly and the lines you did not want gone, sitting where you asked. The translated version carries every one of those decisions rather than starting again from the audio.

Correcting the second language becomes a short list instead of a repeat of the first.

When does this not work well?

A translation that follows your corrected text is exactly what you want for accuracy, and not always what you want for idiom. A line that reads perfectly naturally in the original can come back from Vizard Agent translated literally, and lines like that need a pass of their own.

• Idiomatic lines. Accurate is not always natural.
• Corrections without timestamps. Hard to place reliably.
• A different edit. Timings will not match a recut.
• Very long translations. They may not fit the frame.
• Languages with different word order. Line breaks need re-checking.

How do you fix a result that came back wrong?

Send another short list in exactly the same format as the first one. Vizard Agent keeps the corrected source subtitles and the translation alongside each other, so a fix made in one of them can be carried straight into the other rather than being applied twice by hand.

• "This line is still wrong." Corrected at that timestamp.
• "That line should be gone." Removed rather than reworded.
• "The English reads oddly." Rephrased from the corrected source.
• "They are too long now." Split or shortened to fit.

How does Vizard Agent compare to doing it yourself?

By hand the two languages are separate files corrected separately, so every fix is made twice and they drift. Translating from the audio is also the default in most tools, which quietly discards the corrections you just finished making.

Common questions

Do I need timestamps? They make it reliable. Vizard Agent can find a line without one.

How do I mark a deletion? Say "remove this line". Vizard Agent will not reword it instead.

Will the translation use my corrections? Yes, if it is made after them. Vizard Agent works in that order.

Can I correct the translation too? Yes, in the same format, and Vizard Agent keeps both in step.

Can it fix names? Yes, and Vizard Agent finds names are what automatic subtitles get wrong most.

Where will the subtitles sit? Wherever you say. Vizard Agent moves the whole track.

Can I get the subtitle files? Yes, both languages, from Vizard Agent as files.

What if a translated line is too long? Vizard Agent splits or shortens it to fit the frame.

Does the video get re-rendered? Only if the subtitles are burnt in; Vizard Agent will say which.

Can it do several languages? Yes. Vizard Agent builds all of them from the corrected source.

What about dialect words? Correct them once and Vizard Agent carries them through.

Will the timings change? No, unless Vizard Agent has to split a line.

Can it check the whole track for me? Yes. Ask Vizard Agent to flag lines it is unsure about.

Why not translate from the audio? Because the audio is precisely what produced the mistakes you have just spent an hour correcting, so translating from it hands you the same errors back in a language you may not be able to check them in.