AI Video With Talking Characters

The voice comes out of the same generation as the picture

There are two ways to get a character to talk on screen. One is to render silent video and then drive the mouth with an audio track — the lip-sync route, which needs a face crop, a driving audio file, and produces a face that moves correctly and a body that does not act. The other is to generate picture and sound together, so the performance and the mouth are one decision.

This pipeline does the second. The default video tier is audio-visual: the model is given the line, who says it, how it is said, and a voice reference, and returns a shot with the line spoken in it. Nothing is stitched afterwards.

At a glance

Voices are matched, not synthesised per character

Each character is assigned a voice from a library of pre-built system voices — you hear a three-to-eight-second sample before choosing, can re-match, can pick a different one from the selector, or can upload your own recording. The reason it is matching rather than per-character voice design is cost and stability: designed voices drift between generations and cost per character, while a library voice is the same voice in episode one and episode twelve.

If a character has no voice assigned, the shot falls back to a written description of the voice — age, texture, register — which the model interprets. That is the default for new projects; assigning voices is a deliberate step, not something that happens silently to every character.

The timing rule, and why it exists

A four-second shot cannot hold sixteen words of English. Either the delivery is rushed into unintelligibility or the line is cut. So dialogue shots have a floor derived from the line: word count divided by three for English, character count divided by five for Chinese, and short lines get a pause added on top rather than being crushed to their theoretical minimum. Shots that come back too short for their line are flagged and corrected before anything is rendered — which is cheaper than discovering it in the output.

Language

Dialogue language is a project setting, chosen once, and it is applied where the lines are first written rather than translated afterwards — so the performance is authored in that language instead of being a dub. Fifteen interface languages are available and dialogue is supported across them; a couple of languages occasionally mispronounce individual words on the default model tier, and the selector says so where it is true. The character voices page covers this in detail.

What it will not do

• Clone a real person's voice. Not offered. You can upload your own recording as a reference for your own characters.
• Lip-sync an existing video or photo you upload. Speech comes from generation, not from driving an existing frame.
• Hit an exact runtime for a line. The floor is enforced; the precise delivery length is the model's.
• Burn in subtitles. No text is rendered in frame. Captions go on after export.
• Guarantee perfect pronunciation of every proper noun. Invented names in particular are worth previewing before you commit an episode.

What it costs

Speech is included in the video price — there is no separate audio charge. Video starts at 10 credits per second of output (≈ $0.06/s at the lowest credit price) and runs to $0.47/s on the sharpest tier; assembly is 1 credit per second. Voice matching from the library costs nothing. New accounts get 50 credits and one free preview of up to 5 seconds — enough to hear a character speak before deciding. Rates on the pricing page.

Frequently asked questions

Pricing·Novel to Video·Guides·Samples·Support·Privacy·Terms·Legal·Contact·© 2026 SceneMixer