Start with the result you need
A spoken script and a live conversation solve different problems. For narration, begin with text-to-speech: write a script, generate audio, and download the result. For an assistant that listens and answers, use a speech-to-speech session with microphone input, turn detection, and response playback.
Grok Voice 3.0 is xAI’s conversational speech model, announced on October 14, 2026. Its Voice Agent API is the relevant starting point for real-time conversation. A TTS endpoint is a separate workflow; a successful narration request does not verify a conversational integration.
Try the voice playground
- Return to the homepage playground and write one or two short sentences.
- Select English or Mandarin to match your script.
- Press Generate voice once. The panel shows the queue and generation state.
- Play the audio before downloading it. Listen for names, numbers, pauses, and the last word.
- Rewrite any awkward sentence, then generate a new version.
Keep the first test short. A script containing an introduction, one number, and a question is more informative than a long paragraph. Avoid private customer details in an exploratory demo.
Make a useful listening checklist
Judge the output against the words you submitted. Does it omit anything? Does it pronounce your product name correctly? Is the pace comfortable on headphones and a phone speaker? A natural-sounding first sentence can hide a clipped ending, so listen through to the end.
For a conversational application, add separate tests for interruption, background noise, silence, and a denied microphone permission. Text-to-speech cannot validate those behaviors.
Move from a demo to your application
Create a server endpoint that owns credentials, validates request size, and enforces a usage allowance before contacting a provider. Keep generated results tied to their requesting session. For a browser conversation, issue a short-lived credential instead of placing a permanent API key in JavaScript.
Next, follow the API guide. It separates text narration from real-time audio, so you can build the workflow your users actually need rather than copying an unrelated request example.