OpenAI opens prompt-to-voice to all API customers and adds nine voices to GPT-Live

OpenAI released prompt-to-voice for all API customers; its API changelog lists it on 16 October 2026, against gpt-live-1, v1/audio/voices and v1/live/sessions. You describe a voice in words, and the service creates and saves it for use in GPT-Live. Creating the voice is free; the speech it produces is billed at regular GPT-Live session rates. The same entry adds nine natural voices to GPT-Live: Nuri, Juni and Flitz in German, Brise and Sillage in French, Nira in Hindi, Harema and Shida in Japanese, and Noeul in Korean.

What changed

Prompt-to-voice is generally available to all API customers , and it needs no reference recording and no consent recording. You create a voice through the API or in the voice Playground. Through the API it is a JSON request to/v1/audio/voices with type set to prompt; the guide's example sends aname and the description itself in prompt. An optionalscript_hintguides the generation, and without it the service writes a script from your description ; a script of your own can hold up to 2,000 characters of representative speech, and a blank or too-short one is rejected.

What comes back is metadata — the voice's ID, its creation time and your original prompt — with no audio and no preview. The guide's instruction is to save the returnedidand keep the original prompt. Every creation request saves a new voice , and an organization can create up to 100 prompt-generated voices, a limit that also counts the voices your application creates for its own customers.

To use the voice, set session.audio.output.voice.idto that ID , and put the original description insession.instructions, beside your conversation instructions. The voice is fixed when the session starts; to try another one you open a new session.

Know where the voice stops working. Prompt-generated voices work only with GPT-Live: the Text to Speech, Realtime and Chat Completions APIs do not support them. And creating a voice through/v1/audio/voicesis not compatible with Zero Data Retention, although GPT-Live's speech generations are eligible for it when it is enabled for your organization or project.

The nine new voices sit in the voice table of the live-conversations guide under their API names: nuri, juni, flitz, brise and sillage, thennira, harema, shida and noeul. You pick one by settingaudio.output.voiceto that name when you create the session. If you set none, the default ismarin.

Voices made from recordings stay where they were: limited to eligible customers, who are sent to OpenAI's sales team , at most 20 per organization, with samples of 30 seconds or less.

What it means for you

If you build for Arabic speakers, here is the plain answer: none of the nine new voices is Arabic. Their languages are German, French, Hindi, Japanese and Korean. For voices from recordings, the consent recording must contain one of the listed phrases and nothing else , so look for your language in that list before you plan on a recorded voice. And that same part of the guide says thatgpt-live-1supports custom voices with English accents. The voice table's own advice is to test a voice with the languages and pronunciation your application needs. So an Arabic-sounding voice from a description is something to try, not something the page promises: describe it, run a real conversation in your users' Arabic through it, and judge it by ear before you ship.

When you write the description, use plain language. Accent, age, pitch, timbre and pace are words in your prompt, not separate API parameters. Do not ask for an imitation of a named person, a public figure or a fictional character , and describe an adult: children's voices are not allowed. Change one quality at a time and compare on the same short passage and a few turns of conversation. Rules about answer length or interruptions belong in the session instructions, not in the voice.

Watch the costs that the word "free" hides. Listening in the Playground runs a GPT-Live session and is billed as one. And because every attempt saves a new voice , each wording you try takes one of your organization's 100 places. If your product lets customers create voices, they all share that one quota — decide who may create, and store each ID next to the prompt that made it.

Before you put a voice in front of users, read the conditions. Tell listeners the speech is synthetic, and keep any provenance metadata or watermarks that mark it so. A voice is not guaranteed to be unique: someone else may generate one like it. The terms cover your end users' use, not only yours. And if your organization relies on Zero Data Retention, voice creation falls outside it.

The source

OpenAI's API changelog, the entry of 16 October 2026, with the custom voices guide and the live conversations guide.