Writing Dialogue and Voiceover for AI Video
· 5 min read
Speech-capable models will say whatever you write, or invent something if you don't. A word budget, a voice and one good line.
Some video models, like Veo and Sora, can generate speech and lip-sync it to a character. That is powerful, and it is also where most prompts go wrong. If you describe a voiceover without writing it, the model writes its own words.
Write the exact words, in quotes
Put the line in quotes and say who speaks it: the person on camera, or a voiceover. Keep punctuation simple, and spell out numbers the way you want them said.
Respect the word budget
People speak about 2 to 2.5 words per second. Go over that and the line gets rushed or cut off.
- 4 seconds: about 10 words
- 6 seconds: about 15 words
- 8 seconds: about 20 words
Describe the voice and the room
Add two or three words about the voice, like warm and confident or friendly and upbeat, plus the ambient sound underneath: a quiet café murmur, soft kitchen sounds.
If you don't write the line, the model will. And it rarely says what your brand would.
Dialogue or voiceover?
Dialogue is spoken on camera and lip-synced, best for testimonials and creators. Voiceover is heard over the picture, best for products, places and anything without a face.
A medium static shot of a happy customer in their 30s at a warm rustic kitchen counter. They speak warmly to camera and nod, holding the product. Audio: the subject looks into the lens and says: "Honestly, I wish I had found this a year ago." Natural lip-sync, soft kitchen sounds in the background. Format: 9:16 vertical, 8 seconds.
Tip: Testimonials feel most real with a static camera and natural window light. Let the performance do the work.
Build your prompt in the prompt studio · Check a prompt with the Prompt doctor