Voice
Voice lets your users talk to a Web Agent instead of typing. A microphone button in the chat's composer opens a floating voice bar, and the bar's keyboard button brings the chat back. Both are one conversation.
How it works
The voice runs on OpenAI Realtime, which no other model provider offers. Voice therefore needs an OpenAI API key of its own, whichever model provider the workspace uses. You add it under Settings > Model Provider > Voice. See Add the OpenAI key for voice. What the agent knows and does is the same as in chat: the persona, the knowledge and the actions come from the agent, not from the voice.
The voice bar always floats over the page. The chat closes while the bar is up, and what was said appears in the conversation when the user goes back to it.
Where the agent opens
Opens in decides what the launcher opens.
| Value | What the launcher opens |
|---|---|
| Chat | The chat. The user starts voice from the microphone button. This is the default. |
| Voice | The voice bar. |
When the widget opens by itself on a first visit, it opens the chat. A voice session starts only from the user's own action, because the browser asks for the microphone.
The voice bar
| Shape | What it is |
|---|---|
| Pill | The composer's shape, floating: microphone, waveform, live caption and keyboard button. This is the default. |
| Dock | A dark glass dock with a wide waveform. |
| Orb | An orb that moves with the voice, with a caption and the controls beside it. |
The bar sits at Bottom centre, Bottom right, Bottom left or Top centre. Bottom centre is the default. The bar takes the three Brand colors of the widget.
Voices and languages
| Setting | Values |
|---|---|
| Voice | Marin (default), Cedar, Coral, Sage, Alloy, Ash, Ballad, Echo, Shimmer, Verse |
| Language | Match the user (auto) (default), Spanish, English, Portuguese, French, German, Italian, Dutch, Catalan, Japanese, Korean, Chinese |
With a language selected, the voice always answers in it. With Match the user (auto), it follows the user.
Greeting and style
| Setting | Limit | What it does |
|---|---|---|
| Greeting | 300 characters | With the switch on, the voice speaks first when the bar opens. |
| Voice style | 2000 characters | Instructions for the voice only: tone, formality, phrases to avoid. |
What to know before you turn it on
| Consideration | Description |
|---|---|
| Voice is billed by OpenAI | Voice sessions, and the samples you play in the portal, run on the OpenAI API key you add for voice, and OpenAI bills them to that account. |
| Voice settings belong to a version | They are saved to the draft and reach your users with a publish and a deploy. |
Next steps
- Turn on voice: the steps in the portal.
- Model providers: the keys of your workspace.
- Write the agent's persona: what the agent knows and does, in chat and in voice.
Last updated November 11, 2026