Voice Studio

Voice Studio manages speech profiles, reference samples, generated sounds, and authorized speakers. A Pro GPU Cloud agent deploys the hosted service automatically when that runtime profile is available. Any agent can instead connect to a compatible external Voice Studio over Tailscale.

Create a voice profile

Open Voice and create a profile with a recognizable name. Add a clean reference sample using the capture or upload controls available in the editor.

Use speech with one speaker, little background noise, no music, and consistent distance from the microphone. Only upload or record a voice you have permission to use.

Manage samples

Saved Samples lists references associated with the selected profile. Preview samples before keeping them. Several clean, representative recordings are more useful than a large collection of noisy clips.

Cached sounds are reusable generated assets. They are not training samples unless explicitly added through the profile workflow.

Authorized speakers

Authorized-speaker controls support speaker verification where enabled. Save only consented speakers and remove authorization when access should end. Speaker verification is a product control, not a guarantee against every recording or synthesis attack.

Select the voice

Open Profiles → Voice & Speech, enable Speech Output, select Voice Studio as the engine when needed, and choose the profile. Native audio model providers generate speech directly and ignore the local Voice Studio selection.

External Voice Studio

For a separate GPU computer, join both systems to the same tailnet. During agent onboarding, open Advanced and enter the Voice Studio IP:Port. Enable Use Voice Studio for transcription only when you also want speech-to-text requests routed there.

Data and lifecycle

Conversation speech requests do not need to be added to the Voice Studio generation library. Replayable assistant audio can have a separate retention window. Terminating a hosted GPU agent deletes its local voice profiles and samples, so export what you need first.

For failures, see connection and voice troubleshooting.