How to swap word-by-word captions for readable subtitle lines

Last updated 2026-09-10 · 8 min read

What is the short version?

Word-by-word captions are a social-video convention. On a documentary, a translated interview or anything a viewer is meant to read rather than glance at, they turn the screen into a stutter and make longer sentences impossible to follow.

1. Go to Vizard Agent with the cut.
2. Say it should read like subtitles, not captions.
3. Ask to see one frame before the full render.

What do you need before you start?

The video and a decision about what the text is for. Vizard Agent will produce either style, but they are answers to different questions — one is there to hold attention, the other is there to be read, and the choice decides everything downstream.

• The cut. With or without captions on it already.
• What the text is for. Reading, or rhythm.
• The language. Some scripts force the answer.
• Whether it is a translation. Those need full lines.
• The case. Upper case is a social convention, not a rule.

What do you type into Vizard Agent?

Describe the reading experience to Vizard Agent rather than the mechanism behind it. If you ask for "better captions" what comes back is a restyle of the same thing; asking for lines rather than words changes what the subtitle events underneath actually are.

Prompt

Variants worth knowing:

• "Full-line events." The technical change.
• "Merge into readable units." Not one line per word.
• "Normal sentence case." Upper case is a style choice.

What does Vizard Agent actually do?

Here is the sequence Vizard Agent followed in real sessions where word-level captions had to become proper subtitles. The test render is the important part of it: two frames settle a question that any amount of describing the style will not.

1. Inspects the transcript schema the caption engine uses.
2. Tests a conventional full-line event as a control.
3. Merges the word timings into readable units.
4. Generates the line-oriented track.
5. Test-renders the opening and the ending only.
6. Looks at those frames at full size.
7. Compares a normal-case version against the upper-case one.
8. Checks the line lengths fit the frame.
9. Renders the full video once the style is settled.

Step two is a diagnostic rather than a step towards the result. If a plain full-line event renders correctly, the problem is the word-level reveal; if it does not, the problem is the font or the language shaping and the reveal was a symptom.

Step three is where readability comes from. A subtitle unit is a phrase somebody can take in at a glance, so Vizard Agent merges words up to a sensible length rather than emitting one event per word with the timings unchanged.

Step seven is worth asking for even when you think you know the answer. Upper case reads as social video and sentence case reads as a film, and seeing the same line both ways on a real frame settles a preference that is otherwise argued about in the abstract.

Step five keeps this cheap. Two test frames cost seconds and a full-length subtitle render costs minutes, so Vizard Agent settles the presentation on the opening and the ending before committing to the whole file.

What does the result look like?

Subtitles that sit still long enough to actually be read, break at sensible places in the sentence, and do not flash a new word onto the screen every third of a second. On a translated video the difference is the gap between followable and not followable at all.

Vizard Agent keeps the same word timings underneath the merged units, so the text still tracks what is being said moment to moment.

When does this not work well?

Line-oriented subtitles are the wrong answer for a fast social edit, where the word-by-word reveal is doing real work holding a viewer's attention. They also need considerably more room in the frame, which Vizard Agent will point out is tight in vertical formats.

• Fast social cuts. The reveal earns its place there.
• Vertical video. Full lines take more width.
• Very fast speech. Units get long or change quickly.
• Music videos. Word timing is often the point.
• Mixed styles in one video. Pick one and stay with it.

How do you fix a result that came back wrong?

Say whether the problem is the lines or the letters. A line that is too long and a script that renders wrong look similar at a glance and have entirely different causes, and Vizard Agent will chase the wrong one if the note is vague.

• "The lines are too long." Shorter merged units.
• "The characters are wrong." A font or shaping problem.
• "It changes too fast." Longer units, held longer.
• "I want the reveal back." Vizard Agent restores word events.

How does Vizard Agent compare to doing it yourself?

By hand the caption preset gets chosen once at the start, usually the fashionable one, and the mismatch is noticed on the finished video. Fixing it then means regenerating the whole track, which is why it often does not happen.

Common questions

Which style should I use? Reading means lines. Rhythm means words. Vizard Agent will advise.

Will the timings change? No. Vizard Agent merges the same word timings into units.

Can it do both in one video? Vizard Agent can, but the switch is visible. Usually pick one and stay with it.

Does upper case matter? It is a convention, not a requirement. Vizard Agent will show both.

What about right-to-left scripts? They often need full lines. Vizard Agent tests a control event.

How long should a line be? Short enough to read at a glance. Vizard Agent sizes it to the frame.

Will it break words across lines? No, and Vizard Agent checks the breaks on a frame.

Can I keep the styling? Yes. The event type changes, not the look.

Does it re-render everything? Only after you have approved the test frames Vizard Agent shows you.

What about a subtitle file? Vizard Agent can deliver one alongside the burnt-in track.

Is this the same as a translation? No, but translations almost always want lines.

Can it merge by sentence? Yes, or by a maximum length you give Vizard Agent.

Will it fit vertical video? Vizard Agent will tell you when the lines are too wide for it.

Why did the words look broken? Sometimes it is the reveal; sometimes the font. The control event tells you which.