GPT‑Live‑1 (OpenAI) Voice Model

Overview

GPT‑Live‑1 is OpenAI’s full‑duplex voice model for live conversations. It processes incoming audio while it generates spoken replies, so users can interrupt naturally. It can also hand off deeper reasoning and tool work to a separate backend model, then speak the verified result. This split helps you keep the conversation flowing while longer tasks run in the background.

What you can build

• Customer support phone-style agents that handle interruptions and mid-sentence corrections
• Voice front desks for scheduling, FAQs, and routing to a human when needed
• Language speaking practice that reacts in real time and gives short coaching
• Hands-free field assistants for technicians who can’t use a keyboard
• Live interviewers for screening calls that can ask follow-ups without rigid turn-taking
• Voice companions for onboarding flows and guided product walkthroughs

Inputs

• A fixed voice selection for the whole session. Choose from Marin (default) plus Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder.
• Short conversation instructions in plain text. Use them to set tone, pacing, and when the model should delegate work.
• An optional backend model choice for delegated reasoning and tool use. This backend handles complex steps and returns results for GPT‑Live‑1 to speak.
• An optional reasoning level for the backend work. Higher levels push more deliberation into delegated steps.
• An optional switch to allow the backend to use web search for current facts.
• Optional backend instructions in plain text. Put detailed procedures, business rules, and tool workflows here.
• An audio encoding choice that applies to both user audio input and assistant audio output for the full session.
• A sample rate choice when you use PCM audio. Choose 24 kHz or 16 kHz. G.711 formats always run at 8 kHz.

Outputs

GPT‑Live‑1 returns a live conversation stream with:

• Generated assistant speech as mono audio in your chosen format: - 16-bit little-endian PCM at 24 kHz or 16 kHz, or
  • G.711 μ-law at 8 kHz, or
  • G.711 A-law at 8 kHz.
• Incremental transcripts for both sides of the conversation. You can use them for captions, logs, or UI chat bubbles.
• Delegation-related updates when the model hands off work to a backend, so your app can track what’s in progress.
• Final session closure details, including a final usage snapshot.

Recommended settings

• Use 16-bit PCM at 24 kHz for the best speech quality in most apps.
• Use 16-bit PCM at 16 kHz if you need lower-bandwidth audio and can accept less fidelity.
• Use G.711 μ-law or A-law at 8 kHz for telephony-style pipelines that already run G.711.
• Keep the conversation instructions short. Put long workflows and business rules in the backend instructions.

Limitations

• Model knowledge is frozen at July 31, 2025. Use delegation plus a backend with tools for current facts.
• GPT‑Live‑1 supports audio and text input and output. It does not support image or video.
• The chosen voice is fixed for the full session. Changing voices requires a new session.
• The audio format is fixed for the full session. Input audio is not auto-converted.
• Structured outputs are not supported.
• Fine-tuning is not supported.
• Long conversations may summarize or drop older details. When context usage gets high, the session can switch to a replacement voice engine that receives the original instructions and up to 8,192 tokens of recent history plus summaries.
• Session instructions can be set up to 16,384 tokens.
• OpenAI rate limits for GPT‑Live‑1 are measured in concurrent sessions. Tier 1 allows 25, Tier 2 allows 50, Tier 3 allows 200, Tier 4 allows 300, and Tier 5 allows 500.
• Low-quality audio can hurt results. Background noise, clipping, bad microphones, and overlapping speakers can cause transcript errors and awkward turn-taking.
• Storing a session for later forking depends on storage being enabled. Stored recordings are retained for 30 days. With Zero Data Retention, storage is treated as off and forking is unavailable.

Safety & compliance

OpenAI applies system-level safety checks during the conversation. The system can steer or interrupt a response when it detects risky content. In higher-risk cases, it can end the voice conversation.

GPT‑Live is designed for conversation, not voice impersonation. It uses predefined voices and includes safeguards intended to prevent imitating a real person.

Supported audio generated with GPT‑Live includes SynthID watermarking signals. OpenAI also provides provenance verification tooling so teams can check supported audio for OpenAI signals.