Creativeclaw Clone Voice

Creative Claw — ElevenLabs Voice Cloning

Read shared execution guidance once per task before using tools. It covers existing authorization, model discovery, optional cost checks, imports, and recovery.

Create a reusable voice with ElevenLabs Instant Voice Cloning, attach it to a Creative Claw Character, and use that Character with speech/elevenlabs-v2. Keep this workflow separate from creativeclaw-elevenlabs-v2, which handles steady professional narration and settings. Use creativeclaw-elevenlabs-v3 for expressive audio tags or languages outside v2. Changing the TTS model does not require a new clone and does not upgrade IVC to PVC.

Core workflow

1. Explain that Creative Claw sends the recording to ElevenLabs to create an Instant Voice Clone and stores the resulting voice on a reusable Character.
2. Confirm the speaker owns the voice or has given the user explicit permission to clone and use it. Never infer consent from possession of a recording.
3. Select an existing Character with list_characters, or create one with manage_character({ title, description, image_url? }).
4. Collect a clean 1–2 minute solo recording. Import it into Creative Claw and obtain a durable public audio_url.
5. If the Character already has a voice, explain that cloning again replaces it and obtain confirmation before continuing.
6. Call clone_voice({ character_id, audio_url, consent: true }) only after the explicit confirmation.
7. Generate a short audition with generate_speech({ model: "speech/elevenlabs-v2", character_id, text }).
8. Listen for identity, accent, tone, pacing, noise, pronunciation, and emotional range. Approve the clone before using it for a long script or Film.
9. Use character_id for future speech in that voice. Use creativeclaw-elevenlabs-v2 for detailed script performance, supported pauses, and delivery settings.

Consent and misuse boundary

Proceed only when the user explicitly confirms one of these conditions:

• “This is my voice and I authorize Creative Claw and ElevenLabs to clone it.”
• “The speaker gave me permission to clone and use this voice.”
• The user has a clear license or agreement authorizing voice cloning and synthesis.

Do not set consent: true on the user's behalf without that confirmation. Do not clone a celebrity, politician, public figure, colleague, family member, customer, or any other person merely because a recording is available. Do not help use a clone deceptively, fraudulently, or to imply a real endorsement or statement the speaker did not authorize.

If consent is unclear, pause before the cloning call and ask for a direct confirmation. This pause applies only to cloning; non-mutating recording advice may continue.

Recording the source sample

The clone reproduces what it hears: timbre, accent, cadence, breath, energy, room sound, and artifacts. Recording quality matters more than file size or the number of clips.

Recommended source:

• Sweet spot: 1–2 minutes of continuous clear speech; recommended minimum 1 minute and recommended maximum 3 minutes. These are quality guidelines, not enforced duration limits. Shorter samples can work, but audition them carefully.
• One speaker, one microphone position, one acoustic environment.
• MP3 at 192 kbps or better when possible; a clean WAV is also accepted by Creative Claw.
• No music, other speakers, echo, reverb, fan noise, traffic, clipping, mouth clicks, or heavy processing.
• Consistent volume, distance, tone, accent, and delivery.
• Natural sentences with varied phonemes and punctuation, without long silent gaps.
• A performance that resembles the intended output. Use an energetic sample for an energetic brand voice and a calm sample for a calm narrator.

When meters are available, aim roughly for -23 to -18 dB RMS and a true peak near -3 dB. Do not reject a clean phone recording solely because it lacks studio equipment.

Use isolate_audio only when light cleanup materially improves an otherwise good recording. Preserve the original and audition the cleaned result; aggressive isolation can alter vocal identity.

Create or select the Character

Search first:

list_characters({})

Create a Character when no suitable one exists:

manage_character({
  title: "Maya — Brand Voice",
  description: "Warm, confident founder voice for product narration and launch videos."
})

The Character may also have a visual reference, but an image is not required for voice cloning. Use a clear title because cloned voices remain reusable and must be easy to distinguish.

Import the recording

Use a durable Creative Claw asset URL, not a temporary local path:

• Native ChatGPT attachment: use import_chatgpt_media when exposed.
• Public downloadable URL: use upload_asset.
• Local file in a client that can upload bytes: use get_upload_url, upload the file, then confirm_upload.
• If the client cannot upload directly, open import_media and let the user choose the file.

Do not send private media to an arbitrary third-party URL merely to make it accessible.

Clone the voice

After consent and import:

clone_voice({
  character_id: "<character UUID>",
  audio_url: "<durable Creative Claw audio URL>",
  consent: true
})

This calls ElevenLabs Instant Voice Cloning. It creates a reusable provider voice and attaches its ID to the Character. It is not a one-off reference-audio effect, and it does not train a downloadable standalone model.

clone_voice replaces any existing voice attached to the Character. Treat replacement as destructive: identify the Character, disclose the replacement, and obtain confirmation first.

Audition before production

Use a short script that covers neutral speech, energy, pauses, names, and any target language:

{
  "model": "speech/elevenlabs-v2",
  "character_id": "<character UUID>",
  "text": "Here is a quick voice check. This is the neutral delivery. I look forward to welcoming you to our annual conference.",
  "extras": {
    "voice_settings": {
      "stability": 0.5,
      "similarity_boost": 0.75,
      "speed": 1.0
    }
  }
}

For multilingual use, include names, numbers, and a short sentence in each important language. The source accent normally carries into generated languages, so listen before promising native pronunciation.

Do not pass both character_id and an unrelated curated voice_id. Do not pass the source audio_url to speech/elevenlabs-v2; the reusable Character is the voice selector.

Diagnose a weak clone

If direct ElevenLabs service is unavailable for a private cloned voice, fail closed. Do not silently substitute another provider or curated voice.

Reuse and feedback

For future requests, call generate_speech with the saved character_id, then apply the prompting techniques from creativeclaw-elevenlabs-v2. Use the same Character in Film narration when continuity matters.

Use submit_feedback when cloning fails despite a compliant sample, the result repeatedly misses a specific accent or identity trait, replacement behavior is unclear, or the user requests a stronger cloning model or control. Include the Character ID and concrete quality issue, but do not paste private transcripts or expose the source recording.

Before replacing a good source, audition the existing clone with an appropriate TTS model and settings. V2 does not support v3 square-bracket audio tags; use plain text. Switching to v3 for expression reuses the same Character. Cloning again currently deletes the old voice, so retain the source and do not replace merely to compare v2 and v3.