jambonz Adds Speech-to-Speech Support for xAI's Grok Voice Think Fast 2.0

We’re happy to announce that jambonz supports xAI’s Grok Voice Think Fast 2.0 as a first-class speech-to-speech vendor. If you run jambonz v11 or later with the mediajam media engine, you can connect any phone call to a Grok voice agent with a single verb.

Grok Voice Brings Speech-to-Speech Reasoning to Voice AI Agents

Grok Voice is a native speech-to-speech model: one model listens, reasons, and speaks over a single persistent connection. xAI’s pitch for the 2.0 generation is that it reasons while it talks, so a tool call can fire before the agent has finished its first sentence, instead of after a visible pause, along with better transcription accuracy and a more natural conversational feel than the 1.0 models.

The part we find most interesting for telephony is a setting the other vendors don’t expose: reasoning.effort, which takes high (the default) or none. That is an explicit dial for the tradeoff every voice-agent developer ends up making by hand (deliberation versus latency) and it can be set per session, so a support flow that reasons over tool output and a high-volume “where’s my order” flow can use the same model with different characters.

How to Add Grok Voice to a jambonz Call

If you’ve used the OpenAI Realtime API with jambonz, this will feel instantly familiar. xAI speaks an OpenAI-Realtime dialect, and jambonz exposes it the same way. Here’s a minimal application using the @jambonz/sdk WebSocket interface:

session
  .s2s({
    vendor: 'xai',
    // pinned deliberately, rather than relying on the grok-voice-latest alias
    model: 'grok-voice-think-fast-2.0',
    auth: {
      apiKey: process.env.XAI_API_KEY,
    },
    llmOptions: {
      session_update: {
        instructions: 'You are a friendly and helpful voice assistant. ' +
          'Keep your responses concise and conversational.',
        turn_detection: { type: 'server_vad' },
        audio: {
          output: { voice: 'eve' },
        },
        // xAI-specific: 'high' reasons before speaking, 'none' minimizes latency
        reasoning: { effort: 'high' },
      },
      // the agent speaks first
      response_create: {
        instructions: 'Greet the caller warmly and ask how you can help.',
      },
    },
    actionHook: '/s2s-complete',
  })
  .send();

A few practical notes:

• Pin the model. Name grok-voice-think-fast-2.0 explicitly rather than leaning on the grok-voice-latest alias. That keeps your agent’s behavior from shifting the next time xAI repoints the alias at a new generation.
• session_update is required. Unlike OpenAI, xAI gates audio until it has received your first session.update and replied with session.updated, so there is no “connect and start talking” path.
• turn_detection is top-level. OpenAI GA nests it under session_update.audio.input.turn_detection; xAI expects it directly on session_update. Omit it (or set it to null) to take over turn-taking yourself with input_audio_buffer.commit / input_audio_buffer.clear.
• Don’t set audio formats. The wire format to xAI is pcm16 at 24 kHz and jambonz forces it, so anything you declare in session_update.audio.*.format is overridden.

xAI Grok Voice Documentation and Resources

• The xAI Grok Voice tutorial walks through the full setup, including tool calling and events.
• A complete working example lives in jambonz/v10-examples.
• The llm verb reference documents every xAI option, including voices, input transcription, and tool-call routing.
• xAI Voice Agent support ships in jambonz v11’s feature-server and mediajam.

As always, come find us in the jambonz community with questions. We’d love to hear what you build with it.