Dialog & Lip-Sync
Make AI characters speak convincingly: writing lines that fit the clip, directing voice and tone, and the rules of believable lip-sync.
A character who speaks is worth ten who pose. Talking-head clips are the engine of ads, explainers and social content — and since Veo 3.1 and Sora 2 generate voice and lip movement together with the picture, a one-line prompt can produce a person who says your words to camera. The difference between uncanny and convincing is technique, and it fits in one lesson.
The Golden Rules of Generated Dialog
- Quote the exact line: put the words in quotation marks — do not summarize. She talks about the sale gives you mumbling; she says: followed by the actual sentence gives you the sentence
- Match length to duration: people speak 2-3 words per second, so an 8-second clip fits roughly 12-18 spoken words with natural pauses — write 15, not 40
- One speaker at a time: two characters can exchange short lines, but simultaneous or fast overlapping speech breaks the sync
- Direct the voice: specify tone, pace and accent — says warmly, in a low voice, with a slight French accent — or the model picks a generic announcer
- Ban the captions: end with no subtitles, no on-screen text (Veo especially loves burning them in)
Directing the Performance
Dialog is more than words — direct it like a scene. Describe the physical performance around the line: she pauses, looks down, then delivers the line; he grins before speaking. Emotional stage directions translate into facial acting, and a line delivered after a beat of hesitation reads as human. This is where your camera language pairs beautifully with speech: a slow dolly-in during a confession is a film trope because it works.
A directed dialog shot
Close-up of a young chef in a busy kitchen, wiping his hands on his apron. He looks up at the camera, grins, and says with quiet pride: "Grandma said it would never work. Sold out by noon." Warm tungsten kitchen light, shallow depth of field, soft kitchen clatter in the background. No subtitles, no on-screen text.
Thirteen spoken words for an 8-second clip, a physical beat before the line, tone direction, and the caption ban. Steal this structure.
Lip-Sync Beyond Generation
There is a second path: take an existing clip and re-voice it. Dedicated lip-sync tools map the phonemes of a new audio track onto mouth shapes (visemes) frame by frame — the technology behind AI dubbing that translates a speaker into another language with matching lips. In VAR2, dialog-driven video tools live alongside the generators, and text-to-speech from the audio path covers narration when the speaker stays off-screen — often the simplest fix of all.
You can now make people talk. One lesson left: troubleshooting — because sooner or later a hand will melt, and you will want to know exactly why.
Write a 15-word line, direct the tone, and hear your character deliver it. Create a talking scene