Talk using live voice

Purpose

Mina is cooking with flour on her hands and cannot type, but she still wants to know how much rice to cook for five. Talk to the agent out loud and hear it answer. Your words and its replies are also saved in the chat as text.

Before you start

Use a browser that can record and play sound, and allow microphone access when asked. Live voice is available in Chat mode with the default agent or a custom agent, and in Work mode, provided an allowed voice model is available. To change the voice or how the text appears, see Voice settings.

Steps

1. In an empty message box, select the round button "Live voice chat" at the end of the box. On a project's page, the same button starts a new chat in that project, in the mode chosen at the top (Chat or Work), and opens it in live voice. It takes the place of "Send message" while there is no text or file in the box. Wait while the bar says "Connecting…".
2. When the bar says "Start talking", speak. The bar shows what is happening: "Listening…" while it hears you, "Thinking…" while it prepares an answer, "Using tools…" while it uses a tool such as web search or a connected plugin, and "Speaking" while it answers. The voice model's name is shown under the status.
3. To interrupt a reply, just start talking. The saved text of an interrupted reply contains only what you actually heard.
4. Select the microphone button ("Mute microphone") to stop sending your voice, for example when you step away. The bar says "Microphone off". Select it again ("Unmute microphone") to continue.
5. To finish, select the x button ("End voice mode"). The message box comes back, and the chat shows the whole conversation as messages. A new chat is created on your first spoken message and appears in "Chats".

Expected result

You can ask out loud for anything you can ask in a typed chat: a chart, a document, an analysis of your uploaded files, an image or an answer from your knowledge base. The agent says it is working on it, the result appears in the chat as it would for a typed message, and the agent then tells you briefly what it did. Larger tasks can take a minute.

The chat contains what you said and what the agent answered, in order, and you can keep typing in the same chat afterwards. Spoken replies do not have a "Try again" button. The voice session uses its own voice model, which can differ from the model you picked for typed messages.

If you set "What should Agentum call you?" in your profile, live voice uses that name as typed chats do: the agent may call you by it now and then, not in every reply. A changed name applies from the next voice session.

Words in your Dictionary apply to live voice too, so names and special words are heard and written the way you spelled them.

Live voice in Work

In Work, everything you ask for out loud is done by the Work agent of that chat, as if you had typed it: it reads and writes the chat's files and the project's files, follows the project's instructions and uses the same tools, permissions and usage allowance as a typed Work task. Say what you need, for example "Summarise the contract I uploaded into a one-page Word file". The bar shows "Using tools…" while the task runs, which can take a few minutes. The steps, the files it makes and its written answer appear in the chat, the files show under "Outputs" in the files panel, and the agent then tells you briefly what it did.

Talking again does not stop a task that is running; ending voice mode does. You can end live voice at any point and keep typing in the same Work chat, or start live voice again later in it. See Use Work mode.

Troubleshooting