Full-Duplex Voice

Overview

gpt-live-1 is a full-duplex speech-to-speech model. It listens and speaks at the same time, so the caller can interject mid-sentence instead of waiting for a pause. Select it per session; it is not a default.

Full-duplex vs turn-based

Our other speech-to-speech models are turn-based: the caller speaks, the model listens, the model replies. Interruption is handled as an event — the caller starts talking, playback stops, a new turn begins.

gpt-live-1 does not work in turns. Input audio and output audio are two streams on one session timeline, both live for the whole call:

The two-part split

A gpt-live-1 session is two pieces of work running at once.

1. 1

   The live voice model

   Holds the conversation — hears the caller, speaks, handles overlap and backchannel

2. 2

   The backend

   Reasoning, lookups and tool work, running concurrently with speech

3. 3

   Back into the conversation

   The backend's result is injected into the live session and surfaces as speech

The voice model keeps talking while the backend is still working. That is the point of the design — the caller is not left in silence — but it changes what a few familiar signals mean.

What this changes about how you build

• A completed backend response does not mean the caller heard it. Completion says the work finished, not that the audio reached the other end. Treat confirmation as something the caller says back to you.
• Transcript fragments follow audio cadence, not turns. Caller and agent fragments can overlap, and there is no authoritative turn-completed event. Group fragments into turns in your own code, using the timings on each fragment.
• Context is compacted as the call grows. Older detail may be dropped to keep a long session workable. Keep anything authoritative — the order, the booking, the verified identity — in your own application state rather than relying on the session to still remember it an hour in.

Not supported

Function tools on a gpt-live-1 session belong to the backend half of the split, not to the voice model. For the tool surface available on our other voice agents, see Voice Agent Tools.

Concurrency

gpt-live-1 capacity is measured in concurrent sessions, not requests per minute, and the number available to you depends on your plan. A session that is never closed holds its slot until it expires, so close sessions when the call ends rather than letting them lapse — an abandoned session is a slot another call cannot use.

This is why gpt-live-1 is opt-in per session rather than a workspace default.

Voices

Fourteen voices, default marin:

alloy    ash      ballad   coral    echo     sage     shimmer
verse    marin    cedar    juniper  breeze   vale     ember

A voice outside this list falls back to marin rather than failing the call. The voice is fixed for the session — pick it when you create the session.

Pricing

The per-minute rate covers the whole session, including silence and the time the backend spends working. There is no separate per-token rate for the voice half.

Any reasoning model doing backend work is billed separately, at its own published rate — see the model catalogue. Budget for both.

Selecting it

curl -X POST https://api.callmissed.com/v1/voice/sessions \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "system_prompt": "You are a concise support agent.",
    "llm_model": "gpt-live-1",
    "voice": "marin"
  }'

The rest of the session lifecycle is unchanged — see the Voice Session API. Because gpt-live-1 is one model doing speech in and speech out, no separate speech-recognition or text-to-speech model is selected for the call.