Chat and voice conversations

Chat and voice share one conversation. You can type a follow-up to spoken input, speak after a typed request, and watch tool or background activity without leaving the transcript.

Type a request

Use the composer for an outcome-focused request. Include the target, constraints, and desired result when they matter. Attach supported media when the agent needs to inspect an artifact rather than a description of it.

The send control can become unavailable while the connection is being established. Wait for the workspace connection indicator rather than repeatedly refreshing.

Push to talk

In Profiles → Activation, select Push to Talk. Hold or activate the microphone control, speak, and release when finished. The workspace sends captured speech through the configured transcription path and appends the resulting turn to the same conversation.

Use headphones when speaker audio is likely to feed back into the microphone. Echo cancellation can help, but it cannot correct every room or device configuration.

Wake mode

Select Wake Mode, enter one or more Wake Name(s), and set the follow-up window. After the wake name activates the agent, follow-up speech can continue for the configured number of seconds without repeating the name.

Choose distinctive names and keep the follow-up window only as long as needed. Background speech can otherwise create unintended turns.

Speech output

Profiles → Voice & Speech → Speech Output controls whether assistant responses are spoken. Native audio model providers use their own voice selection. Other providers can use Voice Studio or another configured TTS engine.

Assistant bubbles can retain replayable audio for a limited period. Replay storage is separate from the text transcript and may expire even when the message remains visible.

Follow work in progress

The workspace reports connection state, model work, tool activity, approvals, speech generation, and background-agent progress. A task can continue after the main assistant acknowledges it. Open the activity detail instead of sending duplicate requests.

Interrupt or redirect

Use the stop control to interrupt active speech or agent execution when available. Send a concise steering message to change direction while preserving the current task context. For background work, use Agents Studio to inspect or stop the specific task.

If audio input, transcription, or playback fails, follow connection and voice troubleshooting.