Skip to main content
Voice conversations let users talk to the assistant instead of typing. Click Talk in the message composer to start a live call: you speak, the assistant answers out loud, and the exchange appears in the chat thread as it happens, like a typed conversation. The Talk button appears when voice is enabled for your organization and the selected agent has voice turned on. Contact AGO support to enable voice for your organization.

How It Works

  1. Click Talk in the message composer. The first time, a dialog explains that your voice is streamed to the AI provider and transcribed into the chat. Click Allow microphone to continue.
  2. A voice bar appears above the composer with the call status (“Listening…”, “One moment…”, “Speaking…”) and a microphone level meter, so you know you are being heard.
  3. Your words and the assistant’s replies appear in the thread as you talk. With GPT Live 1, both speakers’ captions can update at the same time and are saved in short groups. With Realtime models, your turn appears after you finish speaking.
  4. Use the microphone button to mute or unmute yourself, and the hang-up button (or Escape) to end the call. The text input is disabled during the call.
  5. When the call ends, the transcript stays in the conversation: you can keep typing in the same thread, share it, or find it later in your history.
You can interrupt the assistant at any time by speaking.

Set Up Voice on an Agent

In AI Studio → Agents, open the agent and, in its general settings, turn on Voice agent. You can then choose:
  • Voice: the voice the assistant speaks with.
  • Realtime model: the model that runs the call. Deployment default uses AGO’s default voice model.
  • Voice style (optional): how the agent should sound when speaking, on top of its persona (for example “Warm and concise. Confirm details before acting.”).
Publish the agent, then start a new call to use the new settings. GPT Live 1 can listen while it speaks and keep the conversation going while the main assistant works on a request. You can add details or correct your request without waiting for the assistant to finish. Its factual answers and actions still come from your agent’s knowledge and tools. An interruption may take a moment to stop its speech. These settings apply to voice calls in the chat. Phone calls to a phone number have their own configuration.

The Conversation Record

  • Saved like any conversation: every spoken turn is kept in the thread and visible in conversation history and analytics. “Voice call started” and “Voice call ended” events mark the call, with its duration and number of turns.
  • The full answer, not just the spoken summary: when the voice assistant consults the main assistant, the consultation appears under the spoken reply as an expandable card with the question asked and the complete written answer.
  • Transcription marker: voice turns carry a small microphone badge. Hover it to see that the text was transcribed automatically and may contain errors.
  • Interruptions: with Realtime models, replies cut off by an interruption are marked Interrupted. GPT Live 1 saves the transcript in short groups and doesn’t mark overlapping speech as interrupted. A transcript records what the assistant generated, which may differ from what finished playing before you ended the call.
  • Save failures are visible: if a turn could not be saved, it’s flagged Not saved in the thread, and the call-ended event shows how many turns were affected.

Requirements

  • A secure (HTTPS) connection.
  • Microphone permission in the browser.