OpenAI lets you generate a GPT-Live voice from a text description

OpenAI shipped prompt-to-voice for GPT-Live on 16 October. Instead of picking from a fixed list of preset voices, you describe the voice you want and the API builds it.

A POST to /v1/audio/voices takes a type of "prompt", a name, and a prompt describing accent, pitch, tone and pace. An optional script_hint, up to 2,000 characters of representative speech, steers the result further. The response includes a voice id, which you then pass into a GPT-Live session:

{
  "type": "prompt",
  "name": "Support assistant",
  "prompt": "A calm, reassuring adult voice, American English accent, mid-range pitch, warm tone, steady pace.",
  "model": "auto"
}

{
  "model": "gpt-live-1",
  "audio": { "output": { "voice": { "id": "voice_123" } } }
}

Creating a voice this way is free; speech generation still runs at standard GPT-Live session rates. OpenAI also added nine new built-in voices covering German, French, Hindi, Japanese and Korean.

Why it matters: teams building voice agents have mostly been stuck choosing from a short list of preset voices or paying to train a custom one. Describing the voice you want in a sentence and getting an ID back removes that step, which matters most if your product needs a specific, consistent voice, a brand’s support line, say, rather than whichever preset sounds closest.