Help center / Guides
Write dialogue for AI video
How to write spoken lines that sound like real people and land on the right speaker: one moment per scene, two to four short lines, one sentence per line.
One scene is one moment
Unless a line says otherwise, these findings come from our own test clips made on Kling 3.0, a model Fantasia does not offer. The steps for Fantasia use its own models, which we have not tested in exactly the same way.
A 10 to 15 second clip holds one moment in one place: two people, a few lines, a few small gestures. Scenes that packed a setup, a gag, a callback and a payoff into 15 seconds felt crammed and jumpy; the one that worked best had four short lines and four small gestures. A longer story is two or three of these moments in a row, each its own clip.
Pick one kind of scene: a talking scene (the lines plus small gestures) or an action scene (one event with a line or two around it), not a full conversation and a big action at once.
Lines that sound like people
- Write the conversation first, two to four lines, then decide the shots around it. Lines slotted into pre-planned shots sounded slotted in.
- Each line answers the one before, right away. It can tease or push back, but it never goes missing.
- Every line makes plain sense if a stranger overheard it. The feeling sits underneath, never in a riddle.
- Talk about a thing in the scene and let the real subject sit under it. At a bowling lane, after a gutter ball: “The floor’s tilted.” “I think it’s just you, watch me.”
- Big moments get the plainest words: “Grandpa, I got one!”
- Nobody explains what the picture already shows.
An exchange in this spirit, at a noodle stall: a courier, wiping his phone on his wet sleeve, “Do you have a charger, any kind?” The cook, still tossing noodles: “Do I look like a phone shop?” Courier: “It’s at four percent, if it dies the app thinks I quit.” She points the ladle at the socket: “Sit there, don’t drip on my stools.”
Getting the right person to say the right line
- Keep each spoken line to one sentence, joining fragments with commas. Written as “Grandma. Look. Perfect.”, the model gave “Perfect.” to the grandmother, and every later line shifted onto the wrong person. Rewritten as “Grandma, look, it’s perfect.”, and with the speaker put on screen first, all four lines landed right. In another scene written the same way, a line still went to the wrong person.
- Before the other person speaks, say who is on screen: “Immediately, close-up of the grandmother, she says, …”.
- Keep speech to about 2 words a second as a rule of thumb, and less in tender scenes.
- Describe how a line is said (“not looking up”, “trying to sound casual”), not an emotion word.
- The speaker faces the camera or is in three-quarter view, in a medium shot or closer. A medium two-shot with both people worked well for back-and-forth lines.
On Kling 3.0, mouths sometimes did not match the words, and the cause looks like the model rather than the prompt. For talking scenes in Fantasia, try Kling 4.0 Flash or Kling 4.0; in one side-by-side test, Kling 4.0 Flash’s lip sync looked fine.
Questions
How many words of dialogue fit in a 10 second AI video?
About 20 is a good ceiling, at roughly 2 words a second, and fewer for a quiet scene.
Why does the AI give my character’s line to the other person?
Lines broken into several short sentences got split between speakers in our tests. Writing each line as one sentence and putting the speaker on screen before they talk helped, but did not always work.
Related
Updated November 18, 2026.