OpenAI lets developers create GPT-Live voices from a text description

What OpenAI released

OpenAI's API changelog for 16 October 2026 lists two GPT-Live updates. Prompt-to-voice is now released for all API customers, letting developers create a custom voice from a text description through the API or the voice Playground and reuse it in GPT-Live sessions. And nine natural voices were added to GPT-Live: Nuri, Juni and Flitz for German, Brise and Sillage for French, Nira for Hindi, Harema and Shida for Japanese, and Noeul for Korean.

GPT-Live-1 is OpenAI's voice model for continuous conversation, able to listen and speak at the same time while a backend model handles reasoning and tools. A day earlier OpenAI rolled out GPT-Live-1.1 and new voices inside ChatGPT. The 16 October changes are for developers building their own voice agents on the API.

How prompt-to-voice works

The custom voices guide describes the flow. A developer sends a JSON request to the /v1/audio/voices endpoint with type set to prompt, a name, and a plain-language description, for example a calm, reassuring adult voice for a customer support assistant with an American English accent, mid-range pitch, warm tone and steady pace. An optional script_hint field takes up to 2,000 characters of sample speech to guide generation; if it is left out, the service writes a script from the description.

Each successful request saves a new voice and returns metadata with an ID, a creation time and the original prompt. It does not return audio or a preview. To hear the voice, the developer starts a GPT-Live session with the voice ID in the audio output settings. OpenAI recommends also pasting the original description into the session instructions so the model performs the voice as intended. The voice is fixed at session start; trying a different one means starting a new session.

No reference recording or consent recording is needed for this route, which is the main change from the custom-voice option OpenAI already offered. The guide suggests changing one quality at a time and comparing results on the same short passage.

What you can describe, and what you cannot

OpenAI lists the qualities a description can cover, as words in the prompt rather than separate API parameters: accent and its strength, perceived age, presentation, pitch, timbre, texture such as raspiness, loudness or vocal effort, speech rate, and how distinctive the voice should be. Two starting points are given, a calm assistant and a lively guide.

The rules are strict on imitation. The guide says not to request imitation of a named person, public figure or fictional character, and children's voices are not allowed. Developers must tell listeners that speech is AI-generated and keep any provenance metadata, watermarks or other signs of synthetic origin. OpenAI also warns that voices are not guaranteed to be unique, because other customers may generate similar ones.

Price, limits and data retention

Creating a prompt-generated voice is free. Speech from it is billed as a normal GPT-Live session, and listening to a voice in the Playground also runs a billed session. OpenAI's pricing page lists gpt-live-1 at $0.05 per minute, billed per second without rounding up, with backend model and tool use charged separately.

Each organisation can create up to 100 prompt-generated voices, and the guide says that cap also applies when an application creates voices for its own customers. That matters for products that might want to let every user design a voice: 100 is a small pool for a consumer app.

There is a data-handling caveat. Voice creation through /v1/audio/voices, including from recordings, is not compatible with Zero Data Retention. GPT-Live speech generation is eligible for Zero Data Retention when an organisation or project has it enabled, but stored session recordings, recording downloads and session forking are then unavailable.

The nine new natural voices

The nine new voices extend GPT-Live beyond English and Brazilian Portuguese. OpenAI's session guide lists each with its language, presentation and source. All nine are marked natural rather than generated: Nuri and Juni are feminine German voices and Flitz masculine; Brise is feminine French and Sillage masculine; Nira is a feminine Hindi voice; Harema is feminine Japanese and Shida masculine; Noeul is a feminine Korean voice.

They join English voices with Australian, British, Irish, North American, Filipino and Southern US influences, and two Brazilian Portuguese voices. Four of the older voices, Quartz, Beacon, Delta and Cinder, are labelled generated rather than natural in the same table. OpenAI advises testing any voice with the languages and pronunciation an application needs.

Voices from recordings are still gated

The other custom-voice route, cloning from a recording, stays limited to eligible customers through OpenAI's sales team. It needs a consent recording in which the voice actor reads a fixed phrase in one of 16 supported languages, plus a sample of up to 30 seconds. Each organisation can create at most 20 voices from recordings.

Recording-based voices work across GPT-Live, the Text to Speech API, the Realtime API and Chat Completions audio output. Prompt-generated voices work only in GPT-Live; the guide says they are not supported by Text to Speech, Realtime or Chat Completions. A developer who wants one designed voice across every OpenAI audio surface still needs the gated recording route.

How it compares with other voice tools

Voice design from text is becoming a standard feature. Open-weight models such as Fish Audio's Drama 3 can be directed in plain language, and xAI has put Grok Voice 3.0 on its Voice Agent API. OpenAI's version is tied to a conversational model that can be interrupted and can hand off work to a backend agent, which positions it for support lines, tutors and in-app assistants rather than narration.

The consent-free design is the notable trade-off. Because no real voice is cloned, OpenAI can open it to every API customer, and it leans on usage policies, the ban on imitating named people and the disclosure requirement to limit misuse. How well those rules hold up will depend on enforcement, which the guide does not describe in detail.

Limitations and open questions

OpenAI has not published quality evaluations for prompt-to-voice, such as how closely generated voices match their descriptions or how consistent a saved voice sounds across sessions. For recording-based voices the guide says gpt-live-1 supports English accents, to be restated in the session instructions; it does not say how well prompt-generated voices handle other languages. AI Lookout has not created or listened to a prompt-generated voice for this article.

The lack of an audio preview in the creation response means every test costs a billed session, and the 100-voice cap and the Zero Data Retention exclusion will shape which products can use the feature. For developers already on GPT-Live, though, it removes the need for voice actors and consent paperwork to give an agent its own voice.