OpenAI Now Builds a Custom Voice From a Text Description

OpenAI released prompt-to-voice to all API customers on October 16, 2026. Its API changelog describes it plainly: create a custom voice from a text description, through the API or the voice Playground, then reuse it in GPT-Live sessions. Creating a voice is free; speaking with it is billed at normal GPT-Live rates. No voice actor, reference recording or consent recording is needed.

For a business building a phone or support agent, that removes the slowest step in giving the agent a voice of its own. Until now the choice was a stock voice that other products also use, or a recorded custom voice that OpenAI limits to eligible customers. A written description sits in between. The trade-off is that the voice is tied to one OpenAI product, and that limit decides whether it is useful to you.

This post covers what the voice guide and changelog say, read on October 19. If you are still choosing a voice platform, our Grok Voice 3.0 vs GPT-Live-1.1 comparison covers the wider decision.

Two changelog entries landed in two days. On October 15, OpenAI added native integration between GPT-Live and its Decisions API, so a voice agent can trigger a structured decision while the caller is still speaking, or react to what the agent says. On October 16 came prompt-to-voice, plus nine new stock voices for 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.

These are API changes. In ChatGPT, OpenAI rolled out GPT-Live-1.1 and new voices in ChatGPT Voice on October 15, which we covered in the voice comparison. The API model is still gpt-live-1, and that is the model prompt-made voices run on.

Why it matters for builders: a recognisable voice is part of how a caller decides whether to trust an agent, and stock voices are shared across every product that picks them. The failure case is treating this as a branding exercise only. The voice you design has to stay intelligible on a phone line, in a noisy shop, for callers with accents; test it there, not in a quiet office.

OpenAI's custom voices guide sets out the flow. You send a request to the /v1/audio/voices endpoint with the type set to prompt and a description of the voice. You can add an optional script hint, up to 2,000 characters of representative speech, to guide how the voice is generated; if you leave it out, the service writes a script from your description. Blank or very short scripts are rejected.

Each successful request saves a new voice and returns its metadata: an ID, a creation time and your original description. It does not return audio or a preview. To hear the voice, you start a GPT-Live session with the voice ID in the session's audio settings, and the guide recommends including the original description in the session instructions as well, so the model performs the voice as described. The voice is chosen when the session starts; trying a different one means a new session.

The example from OpenAI's guide shows the register it expects: a calm, reassuring adult voice for a customer support assistant, with an American English accent, a mid-range pitch, a smooth, warm tone and a steady, conversational pace. Conversation rules, such as answer length and when to hand off, belong in the session instructions, not in the voice description.

The description is ordinary language, not a list of parameters, but OpenAI's guide names the qualities the model responds to. The table below is drawn from it. The guide's advice for refining a voice is to change one quality at a time and compare results on the same short passage and a few conversational turns.

A brand team already has most of this written down. If your company has a tone-of-voice guide for copy, our guide to extracting a brand voice shows how to turn it into concrete descriptors; the step here is translating written tone into audible qualities. "Confident and friendly" in a copy guide becomes something like "a warm adult voice, mid-range pitch, moderate vocal effort, brisk but easy to follow".

The failure case is describing a person instead of qualities. The guide says not to request imitation of a named person, public figure or fictional character, and that children's voices are not allowed. A description that leans on "sounds like" a celebrity is both against the rules and less useful than the qualities you actually want.

The guide is explicit: prompt-generated voices work with GPT-Live and are not supported by the Text to Speech, Realtime or Chat Completions APIs. Voices made from a recording, by contrast, work in all four. That single line decides the use cases.

What fits: a live phone or web voice agent built on GPT-Live, an in-app assistant that talks back, a kiosk or tutor. What does not fit: generating an audio version of a blog post, voiceover for a video, an IVR prompt rendered once and played back, or an existing agent built on the older Realtime API. All of those use the endpoints the prompt-made voice cannot reach.

The example that catches marketing teams: a brand designs a voice for its new support agent and then asks for the same voice on its explainer videos. Under these rules, it cannot have it from this feature. Either the videos use a different voice, or the brand goes through the recording route, if it is eligible, to get one voice that works everywhere.

Each organisation can create up to 100 prompt-generated voices, and the guide says the limit also applies when an application creates voices for its own customers. Every creation request saves a new voice, including each attempt while you refine one. An agency that iterates ten times per client voice across a handful of clients, all under one OpenAI organisation, can reach the ceiling faster than the number suggests.

Two practical responses. Keep client work in separate OpenAI organisations where your contracts allow it, since the limit is counted per organisation, so one client's experiments do not consume another's quota. And refine on a short passage with one change at a time, as the guide suggests, so each saved attempt teaches you something.

Two more terms matter for client work. OpenAI says voices are not guaranteed to be unique: someone else may generate a similar voice, and ownership of output is governed by your agreement with OpenAI. So a prompt-made voice is a brand asset you use, not one you can stop others from approximating. And voice creation through the voices endpoint is not compatible with Zero Data Retention, although GPT-Live speech itself can be, so organisations with that setting need to check before creating voices.

Creation is free. Use is billed through GPT-Live, which OpenAI's pricing page lists at $0.05 a minute for gpt-live-1, billed per second without rounding up to a whole minute. Backend model and tool usage are charged separately. Listening to a voice in the Playground runs a GPT-Live session, so auditioning voices is billed too.

A worked example of the session line: a support agent taking 1,000 calls a month at an average of four minutes uses 4,000 minutes, or $200 of GPT-Live time, before the backend model and any tools. The custom voice adds nothing to that figure. For how this compares with per-character and per-token voice pricing elsewhere, see our voice AI pricing guide.

OpenAI's responsible-use section for prompt-to-voice is short and worth putting into any client contract. Do not impersonate others or break OpenAI's usage policies. Use adult voices only. Tell listeners the speech is AI-generated, and preserve any provenance metadata, watermarks or other signs of its synthetic origin. OpenAI's service terms apply to you and your end users.

The disclosure rule is the one most often skipped in practice. A phone agent that opens with a warm, human-sounding custom voice and no mention that it is an AI goes against OpenAI's responsible-use rules, and may also conflict with local consumer rules. A one-line opener solves it. The failure case is a design review that removes the disclosure because it "sounds robotic"; the rule does not bend for tone.

Designing a voice from text is not unique to OpenAI. MiniMax's price list, read October 19, offers Voice Design, which generates a voice from a natural-language description for use in its text-to-speech models. It costs $3 per voice, charged the first time the voice is used for synthesis rather than when it is generated, with in-API previews billed at $60 per million characters. The difference that matters is the output: MiniMax's designed voices work in its sync and async text-to-speech endpoints, which cover short conversational replies and long-form audio, while OpenAI's work only inside GPT-Live conversations.

Our GPT-Live-1 API guide covers building the agent itself. If you want help designing and testing a voice agent end to end, from the voice description to call handling, that is part of our AI transformation work.