GPT-Live Voices Guide: Prompt-to-Voice, Custom Voices, Cost

Last updated: October 31, 2026

GPT-Live voices gained two upgrades on October 16, 2026. Prompt-to-voice lets every OpenAI API customer create a saved custom voice from a text description, free of charge, for use in GPT-Live sessions. OpenAI also added nine natural voices covering German, French, Hindi, Japanese and Korean.

This guide is based on OpenAI’s developer changelog, the custom voices guide and the GPT-Live conversations guide. For OpenAI’s newest text model, see our GPT-6.1 Terra guide.

Contents

What is new
Ways to set GPT-Live voices
How to create a voice from a description
Writing a good voice description
Custom voices from recordings
Pricing and limits
Limitations
Frequently asked questions

What is new

The OpenAI changelog entry for October 16, 2026 lists two changes for gpt-live-1, OpenAI’s GPT-Live voice model:

Prompt-to-voice for all API customers. You create a custom voice from a text description through the API or the voice Playground, then reuse it across GPT-Live sessions.
Nine new natural voices. Nuri, Juni and Flitz (German); Brise and Sillage (French); Nira (Hindi); Harema and Shida (Japanese); and Noeul (Korean).

Before this release, a custom voice required a consent recording and a reference sample, and that route is still limited to eligible customers. Prompt-to-voice removes the recording step for anyone building on GPT-Live.

The new voices sit alongside the English options with Australian, British, Irish, North American, Filipino and Southern U.S. influence, plus two Brazilian Portuguese voices, listed in the GPT-Live voice options table.

Ways to set GPT-Live voices

There are now three ways to choose a voice. The table compares them using OpenAI’s documentation.

Voice type	How you create it	Availability	Works with
Named voice (for example quartz or nuri)	Pass the API name as a string in audio.output.voice	All GPT-Live users	GPT-Live
Prompt-to-voice	Text description sent to /v1/audio/voices with type prompt	All API customers, free to create, up to 100 per organization	GPT-Live only
Voice from a recording	Consent recording plus a sample of 60 seconds or less	Eligible customers through OpenAI sales, up to 20 per organization	GPT-Live, Text to Speech, Realtime and Chat Completions audio

Named voices are strings, such as "quartz". Custom voices are objects that hold an ID, such as { "id": "voice_123" }.

How to create a voice from a description

The custom voices guide describes this workflow. It needs no recording and no consent file.

Write a voice description covering qualities such as accent, perceived age, pitch, timbre and pace.
Send a POST request to /v1/audio/voices with type set to prompt, a name, your prompt and model set to auto. You can also use the voice Playground instead.
Optionally add script_hint, an example script of up to 2,000 characters that guides generation. If you leave it out, the service writes one; blank or too-short scripts are rejected.
Save the returned id and keep your original prompt. The response contains metadata only, with no audio preview.
Start a GPT-Live session with model set to gpt-live-1 and audio.output.voice.id set to your voice ID.
Repeat the voice description in instructions next to your normal conversation instructions, as OpenAI recommends for consistent delivery.
Connect over WebRTC or WebSockets. For WebRTC, place the configuration in the JSON session field; for WebSockets, send it in the first session.start event.

Each creation request saves a new voice, so repeated experiments count toward the 100-voice limit. Reuse saved voices rather than creating a fresh one for every test.

Writing a good voice description

OpenAI suggests describing the qualities themselves in plain language. Useful dimensions include accent strength, perceived age, presentation, pitch, timbre, texture such as raspiness, loudness, speech rate and how distinctive the voice should be.

A pattern taken from OpenAI’s own example works well as a starting point:

A calm, reassuring adult voice for a customer support assistant. Speak with an American English accent, a mid-range pitch, a smooth, warm tone, and a steady, conversational pace.

Practical tips from the guide:

Change one quality at a time and compare results using the same short passage and a few conversational turns.
Keep conversation rules, such as answer length and interruptions, in the session instructions rather than the voice description.
Describe an adult voice; children’s voices are not allowed.
Do not ask the model to imitate a named person, public figure or fictional character.

If you are comparing voice platforms, our guides to Grok Voice 3.0, Mistral Voxtral 3 and Qwen-Audio 3.1 cover competing realtime and text-to-speech options.

Custom voices from recordings

The older recording route still exists for eligible customers who contact OpenAI sales. It requires a consent recording in which the speaker reads an exact consent phrase, plus a sample recording of 60 seconds or less from the same person.

For GPT-Live, OpenAI recommends a 10 to 60 second sample with several complete sentences, at least five seconds of actual speech, and uploads of no more than 10 MiB. The gpt-live-1 model supports recorded custom voices with English accents; you state the accent in the session instructions, for example “Speak British English.”

The main advantage of this route is reach: recorded voices also work with the Text to Speech API, Chat Completions audio output and the Realtime API, where OpenAI’s example uses gpt-realtime-2.1-mini.

Pricing and limits

According to the custom voices guide, creating a prompt-generated voice is free, and speech in GPT-Live is billed at regular session rates. OpenAI’s API pricing page lists gpt-live-1 at $0.05 per minute, billed per second without rounding up to a whole minute. Backend model and tool usage is charged separately.

Listening to a voice in the Playground also runs a billed GPT-Live session. Other limits from the documentation:

Up to 100 prompt-generated voices per organization, including voices your app creates for customers.
Up to 20 recorded voices per organization.
Voice creation through /v1/audio/voices is not compatible with Zero Data Retention, although GPT-Live speech generation is eligible when Zero Data Retention is enabled.

For how OpenAI handles stored data in its consumer products, see whether ChatGPT trains on your data. For OpenAI’s terminal coding agent, see our Codex CLI 0.166 guide.

Limitations

Prompt-generated voices do not work with the Text to Speech, Realtime or Chat Completions APIs.
The voice is fixed at session start; you cannot switch it mid-session.
OpenAI states that voices are not guaranteed to be unique, so others may generate similar voices.
Creation returns no audio preview, so you must test each voice in a session.
You must tell listeners that the speech is AI-generated and keep any provenance indicators.

Frequently asked questions

What is prompt-to-voice in GPT-Live?

Prompt-to-voice is an OpenAI API feature that creates and saves a custom voice from a plain-language text description. You do not need a reference or consent recording. OpenAI released it to all API customers on October 16, 2026, and the resulting voices work with GPT-Live sessions.

How much does it cost to create a GPT-Live voice?

Creating a prompt-generated voice is free. Speech generated with that voice in GPT-Live is billed at regular session rates, which OpenAI lists at $0.05 per minute for gpt-live-1, billed per second. Backend model and tool usage is billed separately.

Can I use a prompt-generated voice with the Text to Speech or Realtime API?

No. OpenAI says prompt-generated voices work only with GPT-Live. Voices created from an approved recording can also be used with the Text to Speech API, the Realtime API and Chat Completions with audio output.

How many custom voices can an organization create?

An organization can create up to 100 prompt-generated voices, and that limit also covers voices your application creates for its customers. Voices created from recordings are limited to 20 per organization and require eligible access.

Which new GPT-Live voices were added in October 2026?

OpenAI added nine natural voices on October 16, 2026: Nuri, Juni and Flitz for German; Brise and Sillage for French; Nira for Hindi; Harema and Shida for Japanese; and Noeul for Korean.

Can I change the voice during a GPT-Live session?

No. You choose the voice when the session starts. To try a different saved or named voice, start a new GPT-Live session with that voice.

Sources

Image: OpenAI logo (symbol), public domain, via Wikimedia Commons. Placed on a white background and resized. OpenAI and GPT-Live are trademarks of OpenAI. FreeAnswers is not affiliated with or endorsed by OpenAI.

GPT-Live Voices: Power-User Guide to Prompt-to-Voice, Custom Voices and New Languages