Character: identity, face & voice

The tabs that decide who the character is, what it looks like, how it sounds and how it thinks.

Opening a character gives you a persistent menu of settings down the left. This page covers the ones that define the character itself. The rest — documents, memory, tools, phone, publishing — are covered in Knowledge, memory & tools and Faces, phone & publishing.

1 · Identity — who it is

Writing a persona that works

Write it as a description of a person, not a specification of a robot. "You are a helpful assistant" produces a generic answer machine; "You have run the front desk of this hotel for eleven years, you know every regular by name, and you would rather solve a problem than explain a policy" produces a character.

Wake word

Choosing Wake word reveals four extra settings. The character then listens continuously but only responds after hearing one of your phrases — the right choice for a noisy showroom or a device left running all day.

• Wake phrases — one per line. Give people several natural ways to say it.
• Case sensitive — leave off.
• Hide the wake phrase from the LLM — leave on, so "hey alara what time do you close" reaches the AI as "what time do you close".
• Cooldown — the gap before the phrase can trigger again. 1.5 seconds stops a single greeting firing twice.

2 · Avatar Studio — what it looks like

This tab answers two questions: which 3D model is this character, and which shape in that model corresponds to which facial movement.

1. Pick from the catalogue if one is offered — a ready-made avatar arrives fully mapped and you can skip the rest of this tab.
2. Otherwise choose the type that matches where your model came from.
3. Upload the model rather than pasting a URL. Uploading lets the panel read the list of shapes inside the file, which means the dropdowns are filled in for you and typos become impossible. Maximum 100 MB.
4. Press Apply preset if one exists for your type, or Auto-detect to match names automatically.
5. Check the Jaw and Mouth groups. Any name the model does not actually contain is shown in red.
6. Test in Model & Face → Test and raise Global Gain if the mouth movement looks too subtle.

Unreal-side behaviour switches

The bottom of the tab carries switches that the Unreal plugin reads:

• Game Actions — gives the character movement abilities it can decide to use (move to, follow, stop, wait). For game NPCs. Leave off for a kiosk.
• Conversational Gestures — lets the character choose hand gestures while it speaks.
• Teacher Mode — gives it a classroom board it can write formulas and quiz questions on.
• Face Morph Driver — drives facial expression from the emotion engine. Leave this off if your Blueprint already has a face AnimBP, or the two will fight.

5 · Language & Voice — how it sounds

Using your own voice

Free TTS uses ready-made voice models. It does not clone a voice from an uploaded audio sample. To use a custom ElevenLabs voice, add it to your ElevenLabs account, connect your API key, and select it from the voice list.

Choose a speech model, then test the whole reply

The ElevenLabs TTS Model selector offers different balances of response speed and delivery. Start with the panel's conversational default, then compare models using the same voice and sample sentence.

Expressiveness

Four sliders shape delivery: Stability (low is dynamic and emotional, high is even and predictable), Similarity (how tightly a cloned voice is followed), Style Intensity (0 is neutral, 1 is theatrical) and Speaker Boost. The defaults are good; raise Style only if the voice sounds flat.

If the character uses real-time voice

When Core AI is set to Gemini Live or OpenAI Realtime, this tab changes completely: there is no separate speech engine, because the AI produces the audio itself. You simply pick a voice from the grid, filtered by All / Female / Male.

7 · Personality

Six ready-made temperaments — Friendly, Professional, Creative, Witty, Empathetic and Custom — set five traits at once. Pick the one closest to your intent: Friendly for a greeter, Professional for banking or support, Empathetic for care and health contexts.

8 · Core AI — the brain

Choosing a model

Start with a mini or flash class model. They are fast, they are cheap, and for a kiosk or an NPC the quality difference is far smaller than the latency difference. Move up to a flagship model only when you can point at answers that were not good enough.

9 · Guardrails — the rules

Guardrails are instructions about what the character should avoid, written in plain language, one rule per line. When enabled, IAMX includes them in the AI's instructions for standard and native voice conversations. They guide the model; they do not guarantee that every reply will follow every rule.

• Phrase every line as a prohibition: "Do not…"
• Use the Examples buttons for the common ones — political neutrality, no medical advice, never repeat card or ID numbers.
• Keep the list under the character counter shown below the box. If the counter turns red, saving will fail.

10 · Emotion

IAMX tracks a live emotional state for each person the character talks to. After every message the visitor sends, a small classifier nudges eight axes — joy, sadness, trust, disgust, fear, anger, surprise, anticipation — and that state feeds both the character's tone and its facial expression.

Where to go next