# Voice cloning

> Create, preview, and assign an authorized voice sample to an agent.

## Prerequisites

Voice cloning requires an ElevenLabs connection supplied by the deployment or your organization's ElevenLabs TTS configuration. Check availability in the voice library before uploading a sample. You must have permission to use the speaker's voice and affirm consent during creation.

## Create a voice

1. Record one speaker in a quiet room. Use 30 seconds to 3 minutes of clear audio; 1–2 minutes is a useful target.
2. Open the voice library, name the voice, provide the sample, and confirm consent. The upload limit is 20 MB.
3. Wait for creation to finish. If the provider times out, inspect the library before retrying because the provider may have completed the request.
4. Preview the voice with a short, representative sentence.
5. Assign it to the intended agent, test the draft with a Web Call, and publish the reviewed configuration.

The backend normalizes the sample using FFmpeg. Self-hosted operators must have audio processing available in the API runtime.

## Troubleshoot

| Symptom | Next check |
| --- | --- |
| Cloning unavailable | Verify the ElevenLabs connection in Models or ask the operator to enable it |
| Sample rejected | Check file size, supported audio encoding, and duration |
| Another operation is running | Wait for the workspace voice operation to finish |
| Preview differs from a call | Check the agent's effective model configuration and published version |

Remove assignments before deleting a voice that agents depend on. Treat deletion as a provider-backed operation and verify its result in the library.

See [Voice](/configurations/voice) for standard TTS selection and the [API catalog](/api-reference/catalog) for voice-clone endpoints.