OpenAI released prompt-to-voice on October 16, allowing developers to create a voice for GPT-Live from a written description. Access is available to all API customers, and the voice can be created through the testing interface or an API request. According to OpenAI, creating this type of voice is free, while conversations using it are billed at the regular rates for GPT-Live sessions.
The feature is intended for applications using spoken interactions and allows a consistent voice identity to be maintained across different sessions. A developer describes the voice's characteristics, receives an identifier, and subsequently uses it in the application. The documentation for this method states that neither a reference recording nor a consent recording is required. Those materials are needed for a separate workflow based on a person's recorded voice.
The voice description defines its audible characteristics
Characteristics that can be described include accent, the perceived age of an adult voice, pitch, timbre, texture, and speaking pace. They are expressed in ordinary language within the request. OpenAI presents them as qualities to describe, rather than separate numerical parameters for each property. Instructions concerning answer length and conversational behavior belong to the session configuration.
During creation, developers can also provide a short representative script of up to 2,000 characters to guide voice generation. If this is omitted, the service produces a script from the description. Successful requests save a new voice and return metadata, including its identifier, creation time, and original description. The response to the creation request does not include an audio preview file.
The documented limit is 100 prompt-generated voices per organization. The same limit applies when an application creates voices for its own customers. A saved voice identifier can be reused in other sessions without generating another voice. Listening to the result in the Playground interface starts a GPT-Live session, billed under the service's regular conditions.
Compatibility is defined for GPT-Live
Prompt-generated voices work with GPT-Live. The documentation excludes their use in the Text to Speech API, Realtime API, and Chat Completions API. Within a GPT-Live session, the voice is selected at startup, and changing it requires a new session. OpenAI recommends retaining the original description alongside the application's instructions when configuring voice behavior.
GPT-Live manages the spoken conversation, including listening while responding, while a separate backend can perform searches, use tools, and complete tasks. The application establishes permissions and controls operations involving company data. This architecture allows a voice interface to be connected to existing workflows, such as checking an order or looking up a reservation, when the developer has implemented the necessary functions in the backend.
GPT-Live connections can use WebRTC, WebSockets, or telephony and SIP integration. For browser interactions, the documentation requires microphone access, a page served over HTTPS or localhost, and a trusted server holding the API key. The custom voice is selected in the initial session configuration, alongside the model and the mode used to delegate tasks.
In its voice agent documentation, OpenAI separates conversation instructions from the application's operational rules. The voice model receives its role, style, and the conditions for requesting backend help. Work procedures and tools belong to that backend. For a reservation, a spoken confirmation must reflect the outcome of the operation actually performed, while the application remains responsible for authorizing changes.
Interrupting a spoken response and canceling an operation are handled separately in this architecture. A user can continue speaking while the backend is working. The application decides whether an operation continues or is canceled and supplies the result for the spoken response.
Different access conditions for voices created from recordings
The recording-based workflow remains available to eligible customers with the relevant access enabled. It requires two materials: a recording of the person's consent and a sample of the same voice. A voice created through this route can be used in GPT-Live and the other audio interfaces listed by OpenAI. The general availability announced now concerns the description-based method, with its own limits and compatibility.
The prompt-to-voice rules require adult voices, disclosure to listeners that the speech is AI-generated, and preservation of indicators of synthetic provenance. Requests must describe vocal qualities without imitating a named person, public figure, or character. These conditions belong to the voice creation feature and are separate from the commercial rules a company implements in its own assistant.
Voice creation, including creation from recordings, is not compatible with Zero Data Retention. GPT-Live speech generation is eligible for this option when it is enabled for the organization or project. Under Zero Data Retention, stored session recordings, their downloads, and continuation through a fork of a saved session are unavailable, according to OpenAI documentation.
Nine additional voices and session-based billing
The October 16 update also adds nine preset voices: Nuri, Juni, and Flitz for German; Brise and Sillage for French; Nira for Hindi; Harema and Shida for Japanese; and Noeul for Korean. The announced list covers these five languages. Selecting a preset voice and generating one from a description are two separate options when configuring the spoken interface.
GPT-Live bills voice sessions by duration, calculated per second, while backend model and tool usage is billed separately. For applications with AI integration, this separation distinguishes the cost of the conversation from the cost of the operations performed. The new announcement makes creation of a prompt-generated voice free; previewing and using it in conversations remain billable service sessions.
