Make two AI characters sound different

A traveler misses the last ferry. The ticket keeper is still selling tickets. The joke only works if they sound like two people having a conversation.

A recent question about AI characters sounding alike prompted this scene. Cast each role separately, then edit the gaps between their lines. The same approach works for comic animatics, game dialogue tests and storyboards.

This is an animatic: illustrated shots timed to dialogue. Mouths stay still, so you can judge casting and rhythm before committing to character animation.

Cast a contrast you can hear

Write the exchange before choosing voices. Give the characters different intentions as well as different vocal qualities:

Create a short conversation between two characters.

TRAVELER: Bright, brisk and curious. She wants a practical answer.
KEEPER: Lower and unhurried, with dry understatement.
He has accepted the situation before she arrives.

TRAVELER: Has the last ferry gone?
KEEPER: Five minutes ago.
TRAVELER: Then why are you still selling tickets?
KEEPER: Optimism. Terrible business model.

Assign a distinct saved voice to each role. Keep those
assignments consistent through every revision.

Our traveler is Alice and our keeper is George, both listed as British in the voice catalog, recorded with Eleven v4. Briskness and dry understatement were the directions we supplied.

Audition the exchange in context. Can you identify the speaker with your eyes closed? Does the second question react to the answer, or sound like another announcement? If the voices blend together, change the cast before polishing the picture.

Record each role as a complete performance

Record all the traveler's lines together, and all the keeper's lines together. You can then rearrange their turns without changing the source recordings.

We used the Natural preset with one opening direction per role: [bright, brisk, curious] and [lower, unhurried, dry]. Use tags only if your chosen model supports them. Keep stage directions out of the spoken script.

You can do this in a TTS tool and any audio editor. Another option is a dialogue API with explicit text/voice pairs, such as ElevenLabs Text to Dialogue. That is a separate route from the two managed role productions used here.

Direct the silence between lines

Place the role recordings on separate tracks. Split at the natural sentence gaps, preserve breaths where they help, and leave space for a response. Do not stretch the speech to force it into a preset video length.

The keeper's original pause after “Optimism” was 280 milliseconds. We made a second version with an 800-millisecond pause, giving that one-word answer its own beat before “Terrible business model.”

Keep both voices and the traveler's performance unchanged.
Replace the pause after “Optimism.” with exactly 800 ms.
Preserve the original keeper take for comparison.
Do this as a local timing edit, without recording new speech.

That number is a choice for this joke, not a universal dialogue rule. Try the original and edited keeper takes in the download. Listen for whether the extra space creates anticipation or merely makes the reply drag.

Keep the casts and takes together in Voyager

Voyager keeps the saved cast, script, direction and selected take in each role's Voice production. The local pause edit became a separate keeper take, leaving the original available. The traveler's audio stayed unchanged.

We brought the selected takes into an editable Voyager DAW project, then used Voyager Video for the illustration crops and captions. Word timings provide edit points; the DAW leaves the pauses, levels and role tracks available for another pass. Changing one character does not require rebuilding the other character's source performance.

Take the scene apart

Download the scene kit for the role recordings, original and edited keeper takes, captions, illustration, audio stems, an editable Video project, and a script to rebuild the DAW mix. Open index.html locally to play the scene and compare the keeper's timing.

Start with a cast you can recognize without the picture. Then treat pauses as part of the performance. A fraction of a second can change how a line lands, and you can make that decision without buying another voice generation or disturbing the other character.

John Holliman

Cofounder & CTO, Moda

John is the CTO of Moda and a Y Combinator-backed startup founder. Previously CTO of Dover, John builds creative tools that let people direct agents, inspect their work, and refine the result while staying in control.