How to caption a language that does not put spaces between words
Tell Vizard Agent the captions have come back as fragments rather than words. It opens the subtitle file itself to see whether the text was written that way or the renderer is splitting it, rebuilds the captions at phrase level so each glyph sequence stays intact, and checks whether a word tokeniser for that script is even available before trusting any break it makes.
What do you type into Vizard Agent?
Describe what is on screen rather than asking for a style change. A caption track split at the wrong level will survive every style you try, because the styling is applied to whatever chunks the text was already broken into. Say which fault you see, too: one character at a time and letters in the wrong order have different causes.
Prompt
What does Vizard Agent actually do?
Captioning tools break text at spaces. Thai, Lao, Khmer, Japanese and Chinese don't use them, so a tool that expects spaces either makes one caption per character or breaks a word inside a shaped cluster. Both look like a font problem, so Vizard Agent reads the file before blaming the font. In the session behind this article it:
- Inspected the word-timed transcript so each of five vertical Thai clips started on a clean beat, and found an installed typeface with Thai glyph support.
- Generated the caption files, found the first batch had come back empty, fixed the input field and regenerated one file to verify the pipeline.
- Rendered a short test, checked its framing, headline placement and readability, then improved the grouping and tested cleaner one-line chunks at full resolution.
- Opened one subtitle file directly to see whether the renderer was splitting it, and built phrase-level files that kept each glyph sequence complete.
- Checked whether a word tokeniser for the script was installed at all, and when it wasn't, rendered the batch without the faulty caption layer, each clip starting and ending on complete sentences.
When does this not work well?
Word boundaries in these scripts are genuinely hard to find. Without a proper tokeniser for the language, every break Vizard Agent makes is an estimate, and an estimate that lands inside a word is the very fault you came to fix. Mixed scripts in one line, very long phrases, typefaces without full coverage and word-level karaoke highlighting all make it harder.
Common questions
Is it a font problem? Usually not. Vizard Agent reads the file before blaming the font.
What if no tokeniser exists? Vizard Agent delivers without the caption layer rather than guess, and keeps the headline and narration.
Why read the file instead of looking at the video? Because a fragmented file and a renderer that fragments look identical on screen and need opposite fixes.