skills/creativeclaw-clone-voice/references/voices/cloning.md

# Record, upload, clone, audition and reuse

## Recording and consent

Without a consented recording, offer a designed voice instead ([voice design](voice-design.md)).

Ask for at least 60 seconds, preferably 1 to 2 minutes, of natural solo speech. This is Creative Claw's quality recommendation, not a hard minimum shared by all providers. Use a quiet room, a consistent microphone position and a normal speaking voice. Include varied sentences, names and numbers in the intended language. Avoid music, other speakers, echo, clipping, heavy filters and long silences. A clean phone recording is fine.

Explain which provider will process the recording: Cartesia by default, or ElevenLabs when the user chooses it. Obtain explicit confirmation that this is the user's voice or that the speaker authorized cloning and use. Do not infer permission from an uploaded file. Existing explicit consent for the same scoped use need not be requested again. Never set consent true before confirmation.

## Private upload

- ChatGPT attachment: `import_chatgpt_media({ media_file: <actual attachment>, purpose: "voice_clone" })`.
- Interactive recording/file picker: `import_media({ purpose: "voice_clone" })`.
- Local file with byte-upload capability: `get_upload_url({ type: "audio", content_type: <MIME>, purpose: "voice_clone" })`, upload bytes, then `confirm_upload`.

Use the resulting private `audio_asset_id`; do not ask the user to copy an ID already returned. Never publish a private sample to obtain a public URL. A user-supplied direct downloadable `audio_url` is accepted by clone_voice and copied privately, but prefer an already imported asset.

The original is retained privately. Creative Claw prepares a mono MP3 capped at two minutes for the provider. Do not ask the user to re-upload solely because the original is oversized.

## Clone and save

Find an existing Character with `list_characters` when the user names one. Reuse the avatar Character to keep its image and voice together.

```json
{
  "character_name": "My narration voice",
  "audio_asset_id": "<private asset ID>",
  "consent": true,
  "provider": "cartesia"
}
```

Call `clone_voice` with these arguments. For an existing Character, use `character_id` instead of creating a new one. Cartesia is the default provider; pass `provider: "elevenlabs"` for an ElevenLabs clone. The result is saved on the Character automatically, including for a new voice-only Character. No extra save operation is required.

Cloning depends on the workspace's plan and clone capacity. If a tool reports a plan limit, relay its message and link. Replacing a Character's source invalidates its existing provider clones. Identify the Character and obtain explicit replacement direction first.

## Test and reuse

A clone speaks with its provider's model: a Cartesia clone with Cartesia Sonic; an ElevenLabs clone with ElevenLabs v2 for a steady read or v4 for expressive delivery or dialogue. Read [language routing](languages.md) and the selected [Cartesia](cartesia.md), [v2](elevenlabs-v2.md) or [v4](elevenlabs-v4.md) guide. Generate a short audition using the returned character_id, never both character_id and voice_id. Include the user's names, numbers and target-language sentence.

Ask the user to assess identity, accent, clarity and pacing. Do not claim to have listened unless audio inspection was actually available. Adjust text/pronunciation/settings before replacing a good source. Noise or a consistently wrong identity usually merits a cleaner recording, not repeated paid synthesis.

Reuse character_id for future generate_speech calls. ElevenLabs v2 and v4 reuse the same ElevenLabs clone. When the user selects Cartesia, a missing Cartesia clone can be created from the retained consented source. Do not re-clone simply to switch models. Never silently substitute providers or stock voices after a failure.

A Character may contain both images and a voice. Its visual use in generated images/videos does not bring its voice into native video audio. For visible speech, make the speech first and use an audio-driven route such as `video/sync-3`; overlaying speech is not lip-sync. Submit feedback only when the user requests it, without exposing private samples.

Recording guidance: [ElevenLabs instant cloning](https://elevenlabs.io/docs/eleven-creative/voices/voice-cloning/instant-voice-cloning).