OpenAI voices: the full list for GPT-Live, TTS and ChatGPT
OpenAI voices come in four lists, one per product. As of October 16, 2026, GPT-Live defaults to marin and its voice table lists 21 more voices. The Realtime API has 10 voices, and the text-to-speech endpoint has 13. ChatGPT’s app has nine named voices, such as Juniper and Cove, and no API product lists any of them.
We run a phone receptionist on GPT-Live, so we keep track of these lists. The tables below come from OpenAI’s documentation and SDK on that date.
OpenAI voices by product
| Product | Model | Voices | Default or recommended |
|---|---|---|---|
| GPT-Live | gpt-live-1 |
marin plus 21 in the voice table |
marin is the default |
| Realtime API | gpt-realtime models |
10 | OpenAI recommends marin or cedar |
| Text to speech | gpt-4o-mini-tts |
13 | OpenAI recommends marin or cedar |
| Text to speech, older models | tts-1, tts-1-hd |
9 | None stated |
| ChatGPT app | GPT-Live-1 and GPT-Live-1 mini | 9 named voices | You pick one in Settings |
A voice name in one list doesn’t guarantee it works in another. fable, nova and onyx exist only for text to speech, and the 21 GPT-Live voices in the next table exist only in GPT-Live.
GPT-Live voices
GPT-Live is OpenAI’s full-duplex voice model, the one ChatGPT Voice runs on. You set the voice with audio.output.voice when the session starts, and you need a new session to change it. OpenAI’s voice table lists these voices in addition to the default marin:
| Voice | API name | Language | Regional influence | Presentation | Source |
|---|---|---|---|---|---|
| Quartz | quartz |
English | Australian | Feminine | Generated |
| Ripple | ripple |
English | Australian | Masculine | Natural |
| Vesper | vesper |
English | British | Masculine | Natural |
| Willow | willow |
English | Irish | Feminine | Natural |
| Stone | stone |
English | Irish | Masculine | Natural |
| Gleam | gleam |
English | North American | Feminine | Natural |
| Meridian | meridian |
English | North American | Masculine | Natural |
| Bossa | bossa |
Portuguese | Brazilian | Feminine | Natural |
| Tempo | tempo |
Portuguese | Brazilian | Masculine | Natural |
| Beacon | beacon |
English | Filipino | Masculine | Generated |
| Delta | delta |
English | Southern US | Feminine | Generated |
| Cinder | cinder |
English | Southern US | Masculine | Generated |
| Nuri | nuri |
German | Not listed | Feminine | Natural |
| Juni | juni |
German | Not listed | Feminine | Natural |
| Flitz | flitz |
German | Not listed | Masculine | Natural |
| Brise | brise |
French | Not listed | Feminine | Natural |
| Sillage | sillage |
French | Not listed | Masculine | Natural |
| Nira | nira |
Hindi | Not listed | Feminine | Natural |
| Harema | harema |
Japanese | Not listed | Feminine | Natural |
| Shida | shida |
Japanese | Not listed | Masculine | Natural |
| Noeul | noeul |
Korean | Not listed | Feminine | Natural |
OpenAI’s GPT-Live announcement on September 10, 2026 introduced the first 12 rows. OpenAI’s Node SDK (version 7.35.0) also accepts the older Realtime names for Live sessions: alloy, ash, ballad, cedar, coral, echo, sage, shimmer and verse.
The table has no Spanish voice. If your callers speak Spanish, test a few voices with your own greeting before you choose. “Regional influence” describes a voice’s speaking style, so a voice with an Irish influence still speaks English.
GPT-Live sessions cost $0.05 a minute, billed per second, whichever voice you pick.
Realtime API voices
The Realtime API offers 10 built-in voices:
alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin and cedar
You can set the voice when the session starts or on a response. Once the model has spoken in a session, the voice stays fixed. OpenAI recommends marin or cedar for the best quality.
OpenAI text-to-speech voices
The speech endpoint on gpt-4o-mini-tts has 13 built-in voices, per OpenAI’s text-to-speech guide:
alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse, marin and cedar
The older tts-1 and tts-1-hd models support 9 of them: alloy, ash, coral, echo, fable, onyx, nova, sage and shimmer. OpenAI says these voices are optimized for English. You can hear each one on OpenAI.fm, OpenAI’s demo site.
Text to speech reads a script aloud. It can’t hold a conversation, so a phone agent built on it needs separate speech recognition and turn-taking. GPT-Live and the Realtime API handle both.
What are the ChatGPT voice names?
ChatGPT’s app has nine voices. OpenAI’s help center describes them like this:
| ChatGPT voice | OpenAI’s description |
|---|---|
| Arbor | Easygoing and versatile |
| Breeze | Animated and earnest |
| Cove | Composed and direct |
| Ember | Confident and optimistic |
| Juniper | Open and upbeat |
| Maple | Cheerful and candid |
| Sol | Savvy and relaxed |
| Spruce | Calm and affirming |
| Vale | Bright and inquisitive |
You pick one in ChatGPT under Settings, then Voice. Since July 8, 2026, ChatGPT Voice runs on GPT-Live: GPT-Live-1 for paid users and GPT-Live-1 mini for free users.
An earlier voice, Sky, left the lineup in May 2024. Bloomberg reported that OpenAI paused it after listeners compared it to Scarlett Johansson’s voice.
Can you use ChatGPT’s voices in the API?
Not by name. As of October 16, 2026, none of the nine ChatGPT names appears in OpenAI’s GPT-Live voice table, the Realtime list, the text-to-speech list or the Node SDK’s voice type. Some third-party model directories list Juniper or Breeze as GPT-Live API voices. OpenAI’s documentation doesn’t.
You can get the same model. GPT-Live-1 in the API is the model that ChatGPT Voice uses by default for paid users. If you want a receptionist that sounds like ChatGPT, start with gpt-live-1 and try the built-in voices by ear. Our guide on how to build an AI receptionist with the ChatGPT voice covers the phone side.
How do you create a custom OpenAI voice?
OpenAI’s custom voices guide describes two ways:
- From a description. Prompt-to-voice builds a voice from a text prompt. All API customers can use it, creating a voice costs nothing, and the voice works only with GPT-Live. You pay normal session rates when you use it.
- From a recording. You upload a consent recording, in which the voice actor reads a fixed consent phrase, plus a sample. This needs custom-voice API access. The voice works with GPT-Live, text to speech, the Realtime API and Chat Completions audio.
OpenAI’s guidelines ask you not to impersonate anyone, to use adult voices only, and to tell listeners that the speech comes from AI.
Which OpenAI voice should a business phone line use?
Pick by listening, with your own script. Some practical checks:
- Match your callers. A business in Dublin might try
willoworstone. A clinic in Texas might trydeltaorcinder. - Test the greeting. Your business name, street names and service names are where a voice trips. Say them in the test.
- Test on a phone. Call audio sounds thinner than a laptop speaker, so judge the voice over a real call.
- Plan for one voice per call. GPT-Live fixes the voice for the session, so it can’t change mid-call.
LobbyStack, our AI receptionist, runs every call on gpt-live-1 with the default marin voice. If you self-host it, change DEFAULT_VOICE in packages/agent-core/src/live/session.ts to any GPT-Live voice. The code is open source under the MIT license. Hosted accounts use marin; the dashboard has no voice picker today.
OpenAI voice questions
What is OpenAI’s default voice?
marin is the default voice for GPT-Live sessions. For the Realtime API and text to speech, OpenAI recommends marin or cedar for the best quality.
Can you change the voice during a call?
No. GPT-Live needs a new session to change the voice. In the Realtime API, the voice stays fixed once the model has spoken in the session.
Which OpenAI voices speak French or German?
GPT-Live has two French voices, brise and sillage, and three German voices, nuri, juni and flitz. OpenAI says the text-to-speech voices are optimized for English.
Is the ChatGPT voice on TikTok available to developers?
The clips use one of the nine ChatGPT app voices, and the API doesn’t offer those names. Developers get the same GPT-Live model and its own voice list. Our GPT-Live post explains how we connected that model to a receptionist.
Hear a GPT-Live voice on a call
Press the call button on our homepage to talk to LobbyStack’s receptionist on gpt-live-1 and marin. To hear it answer for your own business, create a free account: the Free plan gives you 30 voice minutes a month with no card, and a US or Canadian number free for 14 days. See pricing for paid plans.