Building a voice app on GPT-Live-1? Describe the voice in a prompt, or pick one of nine new language presets

#What changed

OpenAI Developers posted a guide on October 16 for giving GPT-Live-1 apps a voice that fits. There are two new paths. First, prompt-to-voice: you describe the voice you want in ordinary words (accent, age, pitch, tone, pace) and the API saves it as a reusable voice ID. It is generally available to all API customers, and creating a voice costs nothing; you pay normal GPT-Live session rates when you use it. Second, nine new preset voices recorded with native-speaking actors: Nuri, Juni, and Flitz in German, Brise and Sillage in French, Nira in Hindi, Harema and Shida in Japanese, and Noeul in Korean.

This is the developer side of the voice update. The GPT-Live-1.1 voices that landed in ChatGPT, Codex, and dots Voice on October 15 are a separate consumer rollout.

#Who is affected

Developers building voice agents or apps on GPT-Live-1 through the OpenAI API, especially anyone serving users in those five languages. Prompt-made voices only work with GPT-Live. They don't work with the Text to Speech, Realtime, or Chat Completions audio APIs. Cloning a voice from a recording is a separate option limited to eligible customers through OpenAI sales, and it needs a consent recording.

#What to do now

• Try a preset first if one fits: set the session's output voice to its API name, such as noeul for Korean or brise for French, and test it with the real phrases and names your app will say.
• If no preset fits, create a prompt voice in the voice Playground or with a POST to /v1/audio/voices using type prompt. Describe qualities, not people: imitating a named person or character isn't allowed, and voices must be adult.
• Save both the returned voice ID and your original description. The guide says to repeat that description in the session instructions so the voice performs the way you described it.
• Tune by changing one quality at a time and comparing with the same short passage. Each create call saves a new voice, and an organization can hold up to 100 prompt voices, including any your app creates for its own customers.
• Voices are fixed at session start, so open a new session to compare a different one. Tell listeners the speech is AI-generated.

#What is not confirmed

We checked the guide and the voice options table in OpenAI's API docs, but we have not built a test app or judged how the new presets sound. Whether prompt voices will later work outside GPT-Live, and when the GPT-Live-1.1 model itself reaches the API, aren't stated. Voice creation through /v1/audio/voices is not compatible with Zero Data Retention, so check that before using it in a ZDR project.

#Sources

• @OpenAIDevs on X: https://x.com/OpenAIDevs/status/2111142258867335430
• @OpenAIDevs follow-up on the nine presets: https://x.com/OpenAIDevs/status/2111142271177621603
• OpenAI API docs, Custom voices: https://developers.openai.com/api/docs/guides/custom-voices
• OpenAI API docs, GPT-Live voice options: https://developers.openai.com/api/docs/guides/live-conversations#voice-options