elevenlabs-tts

What it does

Treat speech generation as casting, script adaptation, performance direction, rendering, and finishing—not as a single API call. It covers selecting ElevenLabs models and voices, preparing spoken text, controlling pronunciation, pacing, emotion, and code-switching, using TTS APIs, cloning or designing voices with consent, repairing artifacts, mastering deliverables, and evaluating generated speech for rendered speech and voice production.

How it works

• Start by writing a production contract detailing audience, medium, performance direction, duration, latency mode, text volume, identity, pronunciation lists, delivery format, rights, and budget.
• Choose the model by job, using documented model IDs and limits, and test with audition across two plausible model/voice combinations for equal comparison.
• Cast the voice before tuning: select among Voice Library / PVC / Instant Voice Clone / PVC with requirements, auditioning with real script segments to compare intelligibility, pronunciation, credibility, emotion, pacing, and consistency.
• Prepare text for speech separately from the editorial script: resolve ambiguity, expand tokens for spoken form, break dense sentences, use punctuation as performance notation, and mark pronunciation as needed.
• Control pronunciation with a structured ladder: rewrite when needed, try native voice, use alias rules, apply model-supported phonetics, or isolated repairs.
• Manage model-specific pronunciation techniques (IPA in text for v3; SSML-like phoneme tags for Flash; dictionary-driven substitutions for Multilingual v2/Flash v2.5).
• Direct pacing and expression by model: adjust stability, similarity_boost, style, speaker boost, and speed within documented ranges; use breaks only on supported models.
• For Eleven v3, choose stability (Creative/Natural/Robust), use short inline directions, avoid SSML breaks, and generate variants with logs of text, voice, model, settings, seed, and request ID.
• Address multilingual speech and code-switching by using target-language-appropriate voices, localizing pronunciation, and auditing with native listeners; handle language_code usage carefully.
• For long-form narration, decide between ElevenCreative Studio and API-based long-form workflow; split text into paragraphs, maintain coherent chunks, and keep consistent voice, model, and settings across chunks; include context fields when stitching.
• For API behavior, use the complete response endpoint when text is known and playback waits for the full render.

When to use it

Use this skill for rendered speech and voice production tasks including voice selection, text preparation, pronunciation control, and QA of generated speech. It is not intended for music, sound effects, conversational-agent logic, or general transcription.

What it can touch

• Tools: claude-code, codex, copilot, cursor
• It instructs on how to query models, audition voices, and manage text-to-speech rendering workflows; it references using specific ElevenLabs models and endpoints and requires evaluating pronunciation and timing across renders.

Caveats

• It emphasizes that speech production is a multi-step process and warns that longer chunks or misaligned pacing can cause artifacts; it notes model-specific limits and the need to recheck API references and entitlements before production.
• It cautions that some features (like request stitching for v3) may not be available and that legacy model IDs should be treated as migration-only.
• It requires auditing two plausible model/voice pairs and validating against script-specific hardest segments.