How to get subtitles written the way the characters would speak

Last updated 2026-09-22 · 8 min read

What is the short version?

Translation and adaptation are two different jobs. One asks what the words mean; the other asks what this particular character would have said if the scene had been written in your language to begin with, and it is the second one you want from Vizard Agent here.

1. Tell Vizard Agent to adapt for dubbing or subtitling, not translate literally.
2. Say who is young, who is formal, who is deliberately rude.
3. Say the hesitations, insults and exclamations should survive.

What do you need before you start?

The video and a short note on the characters. Vizard Agent hears the performance and can infer a lot from it, but a line about who outranks whom settles the formal and informal forms that a translator would otherwise have to guess at every exchange.

• The video. With the original dialogue.
• The target language. And the register you want.
• Who is who. Ages, relationships, who is rude.
• Subtitles or a dub. It changes the line lengths.
• Anything to keep verbatim. Names, catchphrases, honorifics.

What do you type into Vizard Agent?

Write the rules you would give a human adapter. Terms like "colloquial", "do not translate word for word" and "it has to sound natural spoken aloud" are exactly the instructions a dubbing studio gives, and they work here for the same reason.

Prompt

Variants worth knowing:

• "Adapted for a dub." Names the craft you want.
• "Informal when the character is young." Register, per speaker.
• "Keep the hesitations." They are performance, not noise.

What does Vizard Agent actually do?

Here is the Vizard Agent sequence on an anime clip adapted into French and delivered with burnt-in subtitles. The writing happens against the timings rather than before them, which is the single thing that keeps an adaptation speakable instead of merely accurate.

1. Samples the video and checks what the dialogue transcription offers.
2. Transcribes the dialogue with word timings.
3. Reviews the visuals to see who is speaking and how they are behaving.
4. Writes the timed French adaptation, line by line, against those timings.
5. Uploads the adapted script so it can be checked separately from the video.
6. Checks the expected format for generating the styled subtitles.
7. Verifies the source transcript's timings before building the track.
8. Builds the subtitle track and finds a line offset, where the adaptation and the transcript had drifted apart.
9. Regenerates the track with all thirty-four segments correctly aligned.
10. Renders a short sample first and checks the subtitles are readable.
11. Encodes the full video with the subtitles burnt in.
12. Checks the contact sheet, the waveform and the stream durations before delivery.

Step four is where the difference lives. Writing to the timings means each line has a known number of seconds, so the adaptation is constrained the way a real dub script is — and a line that cannot be said in the time is rewritten rather than squeezed.

Step eight is the failure that always appears in this work. An adaptation merges two short lines or splits a long one, the counts no longer match the transcript, and every subsequent subtitle lands on the wrong speech. Vizard Agent compares the two lists and finds where they first diverge.

Step ten is worth asking for by name. A sample of thirty seconds tells you whether the register is right, and changing the register is cheap before the full encode and expensive after it.

What does the result look like?

Dialogue that sounds like people rather than like a translation: contractions where a person would use them, informal address between friends, insults that land, and pauses left in. The lines fit the time the actors took, and the file is aligned so every subtitle sits on the line it belongs to.

When does this not work well?

Some material resists adaptation. Wordplay that depends on the original language, jokes built on honorifics, or dialogue where the humour is the grammar itself will lose something, and Vizard Agent will tell you what it had to trade rather than quietly flattening it.

• Puns and wordplay. Something has to give.
• Honorific humour. The joke is the form of address.
• Very fast overlapping dialogue. Not enough time for a natural line.
• Songs. Meaning, rhythm and rhyme rarely survive together.
• Cultural references. A local equivalent changes the flavour.

How do you fix a result that came back wrong?

Quote the line and say what is wrong with its voice rather than its meaning. Register notes are easy for Vizard Agent to apply consistently, so a single correction about how one character speaks usually improves every other line that character has in the episode.

• "He would not be this polite." That character's register is lowered throughout.
• "This line is too long to read." It is shortened to the time available.
• "Keep the original catchphrase." That term is left untranslated everywhere.
• "Subtitle 12 is on the wrong line." The alignment is rebuilt from the transcript.

How does Vizard Agent compare to doing it yourself?

By hand you either translate it faithfully and watch the scene die, or you adapt it well and spend an evening per five minutes of dialogue. Fan subtitles are full of both, and the difference between them is not language skill — it is whether someone thought about how the line would be said.

Common questions

Is this a dub or subtitles? Either. Tell Vizard Agent which, because line lengths differ between them.

Can it keep honorifics? Yes. Name the ones to keep and Vizard Agent leaves each of them untranslated.

Will it swear if the original swears? Yes, if you ask for it. By default Vizard Agent matches the strength of the original.

Can I see the script before the video? Yes, and it is the sensible order — Vizard Agent uploads the script separately.

What if two lines overlap? Vizard Agent keeps them as separate timed lines rather than merging them.

Does it handle several characters? Yes, with a register for each one that you can set.

Can it match an existing fan translation's style? Give Vizard Agent a sample and it will follow that voice.

What about on-screen text? That is separate, and Vizard Agent can translate it too if you ask.

Will the subtitles be readable? Vizard Agent checks them on a rendered sample rather than in a preview.

Can it produce an SRT instead of burning them in? Yes, with the same timings.

What if the timings shift? Vizard Agent rebuilds the alignment against the transcript rather than nudging it.

Can it adapt into more than one language? Yes, each with its own register notes.

How long does a clip take? The writing is quick; the alignment checking is what takes the time.

Is a literal version available too? Yes. Ask Vizard Agent for both and compare them side by side.