How to give an AI character a distinct voice
Updated
A character’s voice comes from a few specific choices: sentence rhythm, vocabulary, verbal habits, and how they handle emotion. Two or three sample lines teach those choices to an AI.
You should be able to tell who is speaking without seeing the name. A distinct voice also protects a character from drifting into a generic assistant tone over a long story.
One line, three voices
Here is the same message, “we need to leave now”, from three characters:
- A veteran soldier: “Move. Now.”
- A careful noble: “I’d suggest we take our leave before this becomes unpleasant.”
- A cheerful trickster: “Fun fact: in about thirty seconds, staying here becomes a terrible idea.”
What a voice is made of
Rhythm is the first thing readers notice: short and clipped, or long and winding. Vocabulary comes next, whether formal, slang, old-fashioned, or technical. Then there are habits, such as answering questions with questions, understating everything, or never saying sorry.
A voice also shows in what a character avoids. Someone who changes the subject every time their family comes up tells you a lot without saying it. And it shows in how they handle emotion: some characters make a joke, some go quiet, some get sharp.
Sherlock Holmes has barely shaken Watson’s hand in A Study in Scarlet (1887) before he says “You have been in Afghanistan, I perceive.” One line shows his habit of reading people at a glance.
You don’t need all of these. Two or three strong, specific choices are enough.
Write sample lines
The quickest way to set a voice is to show it. Add two or three short lines, ideally from different situations, so the AI sees the range:
- Greeting someone: “You’re late. Coffee’s cold. Sit.”
- Angry: “Say that again. Slowly.”
- Caught off guard: ”…I didn’t expect you to remember that.”
Keep them short, because the AI tends to copy the length of sample lines too.
Mistakes to avoid
- Overusing a catchphrase. AI models latch onto distinctive phrases and repeat them, so put a signature line in one sample, not in every one.
- Spelling out an accent word by word. It’s hard to read and tends to get exaggerated over time. Suggest the accent through word choice and rhythm instead.
- Describing instead of showing. Replace “speaks in a witty way” with one witty line.
- Similar voices in the same scene. When two characters share a scene, make them contrast: one blunt and one formal, or one talkative and one quiet.
Keeping the voice over a long story
A voice is easiest to keep when it’s set early and corrected as soon as it slips. See how to keep AI characters consistent.
In Lunelle AI you describe a character’s voice to the creator in a chat, sample lines included, and characters are built to keep that voice over long plays. See creating characters by chatting.