Introduction

The API piece of the GPT-Live story just landed. On October 16, OpenAI released prompt-to-voice to all API customers: you describe a voice in plain text, OpenAI generates and saves it, and you use it in GPT-Live sessions. No reference recording and no consent recording needed. Creating the voice is free, and speech is billed at the regular GPT-Live session rate. OpenAI also added nine new preset voices for GPT-Live in five languages. It's closed and API-only, and you can try it today through the API or the voice Playground.

What Shipped

A voice from a sentence

According to OpenAI's API changelog, the Oct 16 update touches gpt-live-1, v1/audio/voices, and v1/live/sessions. The custom voices guide spells out the flow:

  • Send a JSON request to /v1/audio/voices with type set to prompt, a name, your description in prompt, and model: "auto".
  • Optionally add a script_hint of up to 2,000 characters of representative speech. Skip it and the service writes its own sample script.
  • The response gives you a voice id, a creation time, and your original prompt back. It doesn't return audio or a preview, so you listen by running a GPT-Live session.

The limits worth knowing

  • Up to 100 prompt-generated voices per organization. That cap also counts voices your app creates for its own customers, and every successful request saves a new voice, so don't call it on every session.
  • GPT-Live only. Prompt voices don't work in the Text to Speech, Realtime, or Chat Completions APIs. Voices cloned from recordings still cover those, but they stay limited to eligible customers through sales.
  • Free to create, normal price to speak. GPT-Live sessions cost $0.05 per minute, billed per second, according to the Sep 10 changelog entry for the GPT-Live API launch. Listening in the Playground runs a real session and is billed too.
  • Voice creation isn't compatible with Zero Data Retention, though GPT-Live speech itself can be.
  • Adults only, no impressions. Children's voices aren't allowed, you can't ask it to imitate a named person or fictional character, and OpenAI says voices aren't guaranteed to be unique.

Nine new preset voices

If you'd rather pick than design, GPT-Live now has nine natural voices: Nuri, Juni, and Flitz (German), Brise and Sillage (French), Nira (Hindi), Harema and Shida (Japanese), and Noeul (Korean). That's the API side of the accent push we covered in GPT-Live-1.1 in ChatGPT Voice, where the API was still "coming soon."

Try It

Create the voice:

curl --fail-with-body https://api.openai.com/v1/audio/voices \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "prompt",
    "name": "Night radio host",
    "prompt": "A warm adult voice in their 40s with a low-mid pitch and a smooth, slightly husky timbre. American English accent. Relaxed, unhurried pace with gentle vocal effort, like a late-night radio host. Moderately distinctive.",
    "model": "auto"
  }' \
  -o voice.json

Then start a GPT-Live session with the returned ID, and put the same description in your instructions so the performance matches:

{
  "model": "gpt-live-1",
  "instructions": "You are a friendly late-night music guide. A warm adult voice in their 40s with a low-mid pitch and a smooth, slightly husky timbre. American English accent. Relaxed, unhurried pace.",
  "audio": { "output": { "voice": { "id": "voice_123" } } }
}

The guide's table of qualities makes a handy checklist for your prompts: accent, age, presentation, pitch, timbre, texture, loudness, speech rate, and distinctiveness. Describe the qualities themselves, not a celebrity.

My tip, straight from OpenAI's advice: change one quality at a time and compare with the same short passage plus a few back-and-forth turns. The voice is fixed when a session starts, so start a new session to try each version. Since each attempt uses one of your 100 slots, keep a little log of which ID matches which prompt.

Why It Matters for Makers

Until now, a custom OpenAI voice meant a consent recording, a reference sample, and a sales conversation. Now a solo developer can give a voice agent its own character with one paragraph. That echoes the voice design tools ElevenLabs and Gradium already offer, but here the voice plugs straight into a full-duplex conversation model. The trade-off is lock-in: these voices live only inside GPT-Live, so if you need narration files or TTS exports, you still need a different tool. If you're comparing voice agent stacks, our Grok Voice 3.0 piece is the other one to weigh this week.

Conclusion

Prompt-to-voice is small on paper and big in practice: a free, text-described voice for every GPT-Live app, a 100-voice cap per org, and nine new German, French, Hindi, Japanese, and Korean presets alongside it. Write one paragraph about the voice your app deserves, run a session, and tweak one word at a time. I'd love to hear what you end up describing.

Original source:

https://developers.openai.com/api/docs/changelog

Guide: Custom voices · Announcement:

—Anabel ♡