Blog
Updated Guide7 min read

How to Make AI Characters Talk: Voices, Dialogue, and Lip Sync

Give AI characters consistent voices and lip-synced dialogue. Native model audio vs avatar tools vs a voice pipeline, and how to write lines that sound natural.


To make an AI character talk, you need three things: a character who looks the same in every shot, a voice that belongs to that character, and lip sync that matches the mouth to the speech. In Paintbrush, you assign a voice to each character and write their lines in a chapter's script. The chapter is performed as one take with those ElevenLabs voices, and for each shot, Seedance 2.5 hears that recording and performs the on-camera lines itself, voice and mouth together, including on cartoon, anime, and animal faces. There is no separate lip-sync step and no lip-sync charge.

This guide compares the three ways to get talking characters, explains how the Paintbrush pipeline works, and shows how to write dialogue that sounds natural when a machine performs it.

Three ways to make AI characters talk

1. Native audio from a video model

Some video models, including Google Veo, can generate speech along with the clip. That works for a single shot. The catch is that the voice is generated fresh each time, so a character can sound different in the next clip. You also have little control over exactly which line is spoken and how it is timed.

2. Avatar and talking-head tools

Tools like HeyGen and Synthesia animate a presenter reading a script. They are great for training videos and announcements. They are not built for narrative scenes, where characters move through locations and talk to each other.

3. A character voice pipeline

The voice is part of the character, just like the face. The scene is recorded first with each character's voice, and the video is made from that recording, so the delivery and the timing are settled before any video exists. This is the approach Paintbrush uses. The video model still moves the mouth natively, but it performs lines you have already heard and approved, which keeps a cast sounding consistent across a whole video.

How talking characters work in Paintbrush

Assign voices

Pick a voice for each character from the ElevenLabs Voice Library: search voices made for character acting by description (a gravelly villain, a warm grandmother) and filter by gender, age and accent. Each voice has a preview. The creation agent casts from about 50 hand-picked character voices if you would rather not search. Set a narrator voice for the project if your story has narration. Once assigned, a voice follows the character into every chapter.

Write the lines

Dialogue lives in a chapter's script. Each line has a speaker, either a character or the Narrator, and its text, with no quotation marks. Silent beats hold action with no speech. For example:

Narrator: The storm hit just after midnight.
Mara: [whispers] Did you hear that?
Theo: It's only the wind. Go back to sleep.

If a character is talking but not visible, mark the line offscreen. They keep their voice, but no mouth moves on screen.

Add emotion with delivery tags

Speech uses ElevenLabs Eleven v4, which performs audio tags written in square brackets inside a line. Tags are performed, not spoken aloud. Open a line with one, then add a new one where the feeling turns:

[hushed] Eighteenth hole. One putt to win the championship, and the crowd has gone completely silent. [barely audible] He's lining it up now.

Tags can be single emotions like [smug] or [jittery], reactions like [sighs] or [laughs], or short descriptions: [whispering, fearful], [like a sports commentator, speeding up], [starting calm, then losing patience]. Keep them short and describe how the voice sounds. The creation agent tags every line for you, and before recording, new or changed lines are directed again with the whole chapter in view, keeping any tags you wrote.

Timing and shot length

The whole chapter is performed as one take, so every voice hears the others and the pauses between lines come from the performance. The cut director then splits the chapter into shots. Each shot gets the shortest whole-second Seedance 2.5 clip, from 4 to 30 seconds, that holds its slice of the audio, and is trimmed to it, so a line never gets crammed into a rushed clip. A long speech simply becomes several shots, each line staying with its speaker.

Lip sync

For each shot, Seedance 2.5 hears that shot's slice of the chapter recording and performs the on-camera lines itself, voice and mouth together. There is no separate lip-sync pass afterward and no lip-sync charge, in any style and for any number of speakers. Narration and offscreen lines are heard but not mouthed by anyone on screen. This works on stylized, cartoon, 3D, animal, and creature faces as well as realistic ones, as long as the speaker's mouth is visible.

Video models also like to invent voices. So in clips whose lines are all narration or offscreen, Paintbrush removes the generated voices, keeps the ambience, and lays the approved narration back exactly. Clips with on-camera dialogue, and clips with no lines, keep the video model's own soundtrack.

Conversations between two characters

For a back-and-forth, the classic approach is a shot and reverse shot: one shot on the speaker, a cut to the listener's reply. Ask the creation agent for "a shot/reverse-shot conversation between @Mara and @Theo" and the cut director gives each spoken beat its own shot, keeping positions and eyelines consistent. You can also keep both characters in one shared shot when the staging calls for it.

How to write dialogue that sounds natural

  • Keep lines short. One or two sentences per line. Long speeches sound flat and are harder to cut.
  • Write how people talk. Use contractions, fragments, and interruptions. "You're late." is better than "You have arrived later than expected."
  • Use punctuation for pacing. Ellipses add hesitation. Dashes cut a line off. Question marks change intonation.
  • Tag the shifts. A tag carries until the next one, so add a new tag when the feeling changes, not on every phrase. Precise beats broad: [let down] says more than [sad].
  • Contrast your voices. Two similar voices make it hard for viewers to follow who is speaking. Pair a deep, calm voice with a bright, quick one.
  • Leave room for silence. A reaction shot with no dialogue often lands harder than another line.

Troubleshooting talking scenes

The mouth is not moving

The video model can only animate a mouth it can see, so lip sync will not show when the character faces away or something covers the face. Check that the line is not marked offscreen, and describe a framing where the speaker faces the camera, like "medium shot, @Mara facing camera."

The line sounds flat

Make the tags more specific, such as [nervous, trying to sound confident] instead of [nervous], add one where the feeling turns, and rewrite the line the way a person would say it. Shorter lines with clear punctuation usually sound more natural than long sentences.

The wrong character is speaking

Every line in the chapter's script has a speaker. Check the speaker on each line, and give your characters clearly different voices so a mistake is easy to hear.

What it costs

Speech costs 0.008 credits per performed character, or 8 credits per 1,000 characters, and tags and punctuation count. A one-minute conversation of 800 to 1,000 characters costs under 10 credits, and unchanged lines are free when you re-generate. There is no lip-sync charge: Seedance performs the lines as part of the video, which costs 18 credits per second. A 4-second shot with an on-camera line costs 72 credits. A 4-second shot where only the narrator speaks costs 84.5, because it adds 12.5 credits of voice cleanup. Rates are as of November 2026.

Ideas for talking-character videos

FAQ

How do I make an AI character's mouth move with the audio?

In Paintbrush you do not need a separate lip-sync tool. Seedance 2.5 hears each shot's slice of the recorded chapter and performs the on-camera lines itself, voice and mouth together.

Can animals and cartoon characters talk?

Yes. Seedance performs lines on stylized, cartoon, anime, animal, and creature faces as long as the mouth is visible.

Will my character sound the same in every scene?

Yes. The voice is assigned to the character, every chapter is recorded with it, and each shot is performed from that recording, so the voice is not invented fresh per clip. You can recast a voice at any time, and chapters performed afterward use the new one.

Can I use my own voice?

Not currently. Characters use voices from the ElevenLabs Voice Library.

Can two AI characters have a conversation?

Yes. Write each line in the chapter's script under its speaker. The cut director stages the exchange as one shared shot or as a shot and reverse shot per line.

Give your characters a voice.

Create characters, plan scenes, and bring your story to life in one workspace.

Try Paintbrush