Turn on voice

Turn on Voice mode to add a microphone button to the chat and a voice bar to your app. For what voice changes for your users, see Voice.

Before you begin

You have an OpenAI API key. Voice runs on OpenAI Realtime, so it needs one even when your model provider is not OpenAI.

The key is one setting for the whole workspace. Until it is set, the Voice mode switch of every agent is disabled and the row says why.

In the portal, go to Settings and select Model Provider.
In the Voice row, select Edit.
Select OpenAI and enter the key in API key. The key is used only for voice. The portal does not show it again.
Select Continue, then Test connection. The result has a line for voice: Voice key working, or Voice key failed with the reason.
Select Save changes.

The Voice row now shows OpenAI · Realtime and Key configured. A new key applies to voice sessions within five minutes.

To remove the key, edit the Voice row, select Not configured and save. Voice then stops working for every agent in the workspace.

In the portal, go to Web Agents and select your agent.
Select the Look & feel tab and find Voice.
Turn on Voice mode. The rest of the settings appear, and the Live preview switches to the voice bar.
In Opens in, select Chat or Voice. This is what the launcher opens.
Under Voice bar, select Pill, Dock or Orb, then a Position.
In Voice, select a voice. To hear it, select the play button next to the list. The sample is a few seconds of speech in the selected language, billed to the OpenAI account of the voice key.
In Language, select a language, or leave Match the user (auto).
Select Save changes.

Voice is saved to the draft. Your users get it when you publish a version and deploy it.

In the Voice section, turn on Greeting.
Enter what the voice says when the bar opens, up to 300 characters. While the field is empty, Generate with AI writes a greeting from the agent's persona.
Select Save changes.

Symptom	Cause	Fix
The Voice mode switch is disabled	The workspace has no OpenAI key for voice. Voice runs on OpenAI Realtime, which other model providers do not offer.	Add the OpenAI key for voice.
The voice bar opens and the agent does not answer, or a voice sample does not play	OpenAI rejects the voice key, or the account has no access to the Realtime models.	Under Settings > Model Provider, edit the Voice row and run Test connection. Replace the key if the test fails.
Voice is on in the portal and your app has no microphone button	The change is in the draft.	Publish a version and deploy it.