Skip to main content

GPT-Live voice

GPT-Live (gpt-live-1) is an OpenAI voice model for the most natural-sounding calls. You choose it like any other OpenAI voice model: set Provider to OpenAI, then Voice model to gpt-live-1.

GPT-Live listens and speaks at the same time (full duplex), so it handles interruptions, "mm-hmm"s and quick corrections the way a person does. It doesn't wait for a pause, stop and start again.

Behind the voice, GPT-Live uses a reasoning model to work out the caller's request and run the agent's tools, such as a calendar lookup or a booking. The platform chooses and manages this model, so there is nothing to set on the agent. Because the voice keeps the conversation going while that work runs, callers hear "Let me check that for you" instead of silence.

When to choose GPT-Live​

ChooseWhen
OpenAI, gpt-live-1Natural conversation matters most: callers who interrupt, change their mind or talk over the agent, and conversations that mix small talk with tool work. Also for its regional English voices and its Hindi, German, French, Portuguese, Japanese and Korean voices.
OpenAI, other Realtime modelsYou want to tune turn-taking yourself with Barge-in sensitivity and Silence timeout, or choose the transcription model. Also for phone numbers and SIP trunks today.
Google Gemini LiveYou prefer Gemini's voices or models.
ElevenLabsYou want a specific ElevenLabs voice, and the agent's chat model writes the replies.
SarvamYour callers speak Indian languages, and the agent's chat model writes the replies.

For how the providers and models work, see Models and voices.

Where it works​

ChannelGPT-Live
Website widget voiceYes
Playground Talk in browserYes
Phone numbers and SIP trunksNot yet. Use OpenAI Realtime for phone agents for now.
Text chatNot used. Chat runs on the agent's chat model.

Plans and pricing​

GPT-Live is available on the Standard (pay-as-you-go) and Enterprise plans. If your plan doesn't include it, gpt-live-1 shows as locked in the Voice model list and you can't select it. See Usage and billing.

GPT-Live must be turned on before it appears. If gpt-live-1 isn't in the Voice model list at all, it isn't enabled for your workspace yet. Contact support.

GPT-Live calls are billed per minute in their own GPT-Live price tier. You can see it under Usage rates on the billing page. For current prices, see voisx.ai/pricing.

Set up an agent​

  1. Open the agent and choose Model & Voice.
  2. Under Voice, set Provider to OpenAI.
  3. Set Voice model to gpt-live-1.
  4. Choose a Voice persona. The list shows the GPT-Live voices.
  5. Save, then try it in the Playground.

The platform handles transcription for gpt-live-1, so the Transcription model setting doesn't apply.

Voices​

GPT-Live speaks all ten OpenAI Realtime voices, plus its own regional and language voices. The default is marin. The recommended voices are marin and, for Hindi, Nira.

Shared with OpenAI Realtime: alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin, cedar.

GPT-Live only:

VoiceLanguageAccentVoice type
QuartzEnglishAustralianFemale
RippleEnglishAustralianMale
VesperEnglishBritishMale
WillowEnglishIrishFemale
StoneEnglishIrishMale
GleamEnglishNorth AmericanFemale
MeridianEnglishNorth AmericanMale
BeaconEnglishFilipinoMale
DeltaEnglishSouthern USFemale
CinderEnglishSouthern USMale
BossaPortugueseBrazilianFemale
TempoPortugueseBrazilianMale
NuriGermanFemale
JuniGermanFemale
FlitzGermanMale
BriseFrenchFemale
SillageFrenchMale
NiraHindiFemale
HaremaJapaneseFemale
ShidaJapaneseMale
NoeulKoreanFemale

Pick a voice that matches the agent's Language, set in the Overview section. A language voice speaks its language most naturally. Test it with the names and words your callers use.

You can't change the voice during a call. To compare voices, use the Test session settings in the Playground: they set the voice model, voice and language for one test call without changing the agent.

Write prompts for GPT-Live​

You write one system prompt, as for any agent. GPT-Live uses it for both the conversation and the reasoning, and the platform adds the rules each part needs.

  • Keep persona and speaking style short. Start the prompt with a few sentences on who the agent is and how it sounds, for example "You are Maya, a calm, friendly receptionist for Bright Dental. Speak warmly and at an unhurried pace." The voice model follows this on every turn.
  • Business rules and tool steps stay in the main prompt. Opening hours, prices, policies, what to collect and in which order, and when to use each tool go in the system prompt as usual. GPT-Live uses all of it when it reasons and runs tools.
  • Don't script every sentence. Describe the behavior you want and let GPT-Live choose the wording for ordinary replies. Use the greeting settings when the opening words must be exact.
  • Don't write turn-taking rules. Rules such as "never speak while the caller is speaking" can make GPT-Live sound stiff. It already stops when it is interrupted.
  • Tools work as before. Tool settings apply unchanged. A tool set to Ask the caller to confirm before running still reads the details back and waits for an explicit yes before it runs. See Tool settings.

The tips in Voice best practices, for example on emails, phone numbers and time zones, apply to GPT-Live too.

Call settings​

Call settings apply to GPT-Live calls, with two exceptions: Barge-in sensitivity and Silence timeout have no effect, because GPT-Live handles turn-taking itself. Max call duration, hang up after silence, end-call phrases and the recording notice work as usual.

Was this page helpful?