Creative Claw — ElevenLabs v3

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

Use speech/elevenlabs-v3 for expressive acting, audio tags, and languages outside Multilingual v2. Prefer speech/elevenlabs-v2 for steady professional narration and identity-focused clone auditions in supported languages. Choose a voice that already resembles the requested age, energy, accent, and performance; inline tags shape delivery but cannot completely transform an incompatible voice.

Core workflow

1. Identify language, audience, format, speaker profile, energy, pronunciation risks, and target duration.
2. Reuse the requested voice or select a fitting default (Hale only for an unspecified English brief). Present choices only when the user wants to choose or the casting decision is material.
3. Rewrite written copy into natural spoken language only when the user permits copy edits. Preserve exact approved wording otherwise.
4. Add sparse inline [audio tags] at performance changes.
5. Call get_model_params({ model: "speech/elevenlabs-v3" }) and use current voice IDs and fields.
6. Call generate_speech with model: "speech/elevenlabs-v3", the selected voice_id, and the performed script.
7. Listen for pronunciation, emotion, pacing, clipped words, tag leakage, and unwanted accent shifts. Regenerate only the affected segment when possible.
8. Use returned timestamps for subtitles, captions, or lip-sync alignment.

Voice selection