Choose the voice and personality

Updated November 3, 2026, based on the xAI Text to Speech and Speech to Speech pages and voice prompting guide. xAI offers 28 built-in voices, shared by the voice agent and text-to-speech; the descriptions and use cases below are xAI’s own. Sources: Text to Speech · Speech to Speech · Prompting Guide.

Twenty-eight voices, twenty-eight first impressions

The Grok Voice Agent gives you 28 voices (as of November 3, 2026), each with its own timbre and temperament, and every one of them speaks every supported language. This choice is more than cosmetic. Before the first sentence is over, users have already formed an idea of who they’re talking to. An upbeat voice delivering an order cancellation sounds wrong; a deep, measured voice on a product demo puts the room to sleep.

The list changes over time. GET /v1/tts/voices returns the current one, and the Custom Voices API lets you create a voice from a short reference clip; you then use its ID exactly like a built-in voice.

Set the voice

You pick the voice in the voice field of the session.update message. It takes three lines:

{
  "type": "session.update",
  "session": {
    "voice": "rex"
  }
}

The name is case-insensitive: "Eve", "eve", and "EVE" all refer to the same voice. The text-to-speech documentation lists eve as the default voice; for the agent, always set the voice explicitly. As with the rest of the configuration, sending a new session.update mid-conversation changes the voice. A phone reception flow could start with an upbeat greeting voice, then switch to a calmer one as soon as the call turns to a sensitive topic.

Match the timbre to the behavior

The voice only does half the job: it sets the texture, while the instructions set the conduct. A warm voice that rattles off curt answers creates a dissonance users feel without being able to name it. The three configurations below show three consistent pairings of timbre and role.

For customer service, celeste, which xAI lists for support, goes with a patient tone and a way out to a human:

{
  "type": "session.update",
  "session": {
    "voice": "celeste",
    "instructions": "You are a customer service assistant for the Maison Verte store. You are patient, empathetic, and always ready to find a solution. Use simple, reassuring language. If you can't solve a problem, offer to transfer the customer to a human agent."
  }
}

For a technical tutorial, cosmo, listed for education, pairs with instructions about structure and precision:

{
  "type": "session.update",
  "session": {
    "voice": "cosmo",
    "instructions": "You are a technical trainer who specializes in web development. Explain concepts in a structured way, with concrete examples. Be precise and direct. Suggest hands-on exercises when relevant."
  }
}

For a company onboarding program, leo brings an authority that the instructions soften with warmth and regular check-ins:

{
  "type": "session.update",
  "session": {
    "voice": "leo",
    "instructions": "You are the official trainer for the new employee onboarding program. Present information with authority but with warmth. Follow the program in order and check for understanding after each point."
  }
}

How to decide

Four considerations come up in every project. The first is brand fit: a premium brand will see itself in altair, a laid-back startup in castor or iris, and getting it backwards creates an instant mismatch. The second is the context of use, which sometimes outweighs the brand: a wellness assistant calls for a voice xAI lists for wellness, such as carina, lux, or liora, while a phone switchboard does well to reassure callers with celeste or aurora. The third is how long people interact: over long sessions, a calm voice is less tiring than a very distinctive one. The fourth, and the most important, is that how people perceive a voice is subjective. Before you commit, play three candidates to a handful of real users, in your target language. It takes an hour and saves you from discovering after launch that the voice you picked annoys half your customers.

What speech changes in your prompts

Moving to voice breaks some prompt-writing habits and calls for new ones. The xAI voice prompting guide suggests a five-part structure: role and persona, goal, conversation flow, guardrails and escalation, and voice style. For voice style, it gives specific advice. Keep to one or two short sentences per turn unless the caller asks for more detail, and pause briefly after each question. If the caller goes quiet, follow up with a short question. And when what they said was inaudible or incomplete, ask a short clarifying question rather than guess.

Open-ended questions work well in writing but often throw people off in conversation. “Would you like me to check your order?” works better than “What can I do for you?” And an agent that asks “Could you say that again?” beats one that guesses, because in a spoken exchange a misunderstanding takes two turns to fix.

Key takeaways

• 28 built-in voices are available, shared by the voice agent and text-to-speech, plus custom voices
• xAI describes the tone and use cases of each voice: start from those use cases, then test with real listeners
• You can change the voice mid-session without dropping the connection
• Structure instructions the way the xAI guide suggests: one or two short sentences per turn, brief follow-ups, clarifying questions
• Match the voice to your brand, the context of use, and the length of interactions