OpenAI opened voice creation from text descriptions to all API customers on October 16, 2026. For customer-support and product teams, that means a new way to give an assistant its own sound without recording a speaker. The company calls the feature prompt-to-voice. OpenAI’s announcement.
The voices are for GPT-Live, the service that handles spoken conversations inside an application while a separate backend looks up information or completes tasks. This is an API feature for building apps, not a new ChatGPT voice setting. How GPT-Live works.
Which voice route fits your assistant?
Our recommendation is to choose the software first, then the sound. Start with a preset if it fits; use a text description for a more tailored GPT-Live voice; consider the recording route when a particular consenting speaker is essential.
Route | Best starting point | Main boundary |
|---|---|---|
Preset voice | An available voice already suits your audience. | No separate voice-creation step. Check the intended language and pronunciation. Voice options. |
Text-described voice | Your app uses GPT-Live and presets do not fit. | No recordings needed, but the saved voice works only in GPT-Live. Creation reference. |
Voice from a recording | You need a particular consenting speaker’s voice. | Eligible customers only; requires consent and reference recordings. Supports GPT-Live, Text to Speech, Realtime and audio-output Chat Completions. Recording requirements. |
The preset choice also widened on October 16: OpenAI added nine voices across German, French, Hindi, Japanese and Korean. That gives multilingual teams another reason to check the existing options before commissioning a custom voice. The same-day preset release.
A voice created from text cannot simply replace one in an existing Realtime or text-to-speech application. Ask the supplier which voice system the product uses before approving the design. Supported uses of prompt-created voices.
A tailored voice is not an exclusive voice
Describe qualities such as pitch, accent and pace, rather than asking for a named person’s likeness. OpenAI prohibits impersonation and children’s voices, requires disclosure that speech is AI-generated, and requires preservation of indicators of its synthetic origin. Prompt-to-voice conditions.
OpenAI does not guarantee uniqueness: someone else may generate a similar voice. Treat customization as a way to shape presentation, not a promise of an exclusive brand sound. Voice uniqueness limits.
Keep the chosen voice’s saved ID and original description for reuse. The allowance is 100 prompt-generated voices per organization, including voices created for its customers—not 100 per customer. Saving and reusing voices.
Free to create, paid to hear
Creating the voice costs nothing, but listening to it in OpenAI’s Playground runs a billed GPT-Live session. Preview billing.
As of October 16, the public GPT-Live rate is US$0.05 per minute, billed per second without rounding to whole minutes. Backend models and tools cost extra. Current API pricing.
Budget for the whole active session, including silence and time spent waiting for backend work—not just the assistant’s speech. Muting the microphone does not stop billing; the session must be closed. How session duration is charged.
Illustrative audition budget: ten successful sessions with exactly two billable minutes each would cost 10 × 2 × US$0.05 = US$1.00 for the voice component. This assumes no extra sessions or failed starts and excludes backend work, telephony, hosting and taxes. It is an estimate, not a service quote. Rate used in the calculation.
For WebRTC connections, the initial 15-second charge is credited against a running session’s duration; do not add it twice. Session creation and reconnects still need to be accounted for. Initialization billing.
Before you settle on a voice
Compare candidates using the language and pronunciation your application needs. Voice selection is fixed when a GPT-Live session starts, so trying another voice requires a new session. A listed language is not a guarantee that every product name or customer phrase will sound right. Session and voice guidance.
If your business requires OpenAI’s Zero Data Retention controls, review voice creation separately from conversations. The saved voice is retained until deletion, and voice creation is not eligible for those controls. GPT-Live sessions can qualify, but session storage, recording downloads and forking are unavailable under that setting. Data-retention rules.
Finally, keep the voice decision separate from what the assistant is allowed to do. GPT-Live leaves permission checks and required confirmations with the application; a more suitable voice does not change that responsibility. Application responsibilities. For that separate design choice, see Choosing when GPT-Live voice assistants should act.