Elevenlabs Tts

ElevenLabs TTS Skill

Generate high-quality speech for Pompom companion using ElevenLabs Text-to-Speech API v3.

When to Use

• Pompom speaks via the TTS voice system (pompom-voice.ts ElevenLabsEngine)
• Voice test (/pompom:voice test)
• Commentary, assistant replies, and reaction speech events

Pompom Voice Config

• Default Voice: Jessica Anne Bogart - Chatty and Friendly
• Voice ID: g6xIsTj2HwM6VR4iXFCw
• Model: eleven_v3
• API Key: ELEVENLABS_API_KEY from environment (~/.env.secrets)
• Privacy: Zero Retention Mode — enable_logging=false query param + xi-no-log: true header
• Stability: 0.0 (Creative — maximum expressiveness for animated character)
• Similarity Boost: 0.85
• Style: 0.8
• Speed: 1.1 (slightly faster for punchy animated character lines)
• Speaker Boost: true
• Output: PCM 24kHz, wrapped in WAV for playback

Available Voices

Integration with Pompom

The voice engine is in extensions/pompom-voice.ts (ElevenLabsEngine class). Speech events flow:

1. say() in pompom.ts fires a SpeechEvent
2. pompom-extension.ts routes it to enqueueSpeech()
3. Speech queue applies personality gating, cooldown, dedup
4. ElevenLabsEngine.synthesize() calls the API
5. Audio plays at configured volume via native OS player (afplay/paplay/aplay)

Eleven v3 Best Practices

Voice Settings for Animated Characters

• Stability 0.0 (Creative): Required for audio tags. Maximum expressiveness — perfect for animated characters with short, emotional utterances
• Similarity Boost 0.85: High fidelity to source voice while allowing emotional range
• Style 0.8: Strong stylistic expression for character personality
• Speed 1.1: Slightly faster delivery suits short, punchy lines like "Yum!", "Wheee!", "Got it!"

Audio Tags (v3 Only)

Eleven v3 supports emotional audio tags — used throughout Pompom's speech:

• [excited] — energetic reactions (ball, dance, game, treats)
• [happy] — warm moments (petting, hugs, feeding, accessories)
• [whispers] — quiet moments (sleep, nap)
• [sighs] — contentment (waking from nap)
• [mischievously] — playful moments (peekaboo)
• [sings] — musical moments (la la la)
• [curious] — weather observations (clouds rolling in)
• [concerned] — weather warnings (storm brewing)

Rules:

• Tags only work with stability 0.0 — NOT 0.5 or 1.0
• One tag per utterance (Pompom's lines are short)
• Match tag to the voice's character — don't force mismatched emotions
• Tag goes at the start of the text: [excited] Catch the stars!

Text Preparation

• Keep sentences short and conversational (Pompom's lines are 2-6 words)
• Use contractions ("I'm", "that's", "let's")
• Use exclamation marks for energy, ellipses for pauses
• CAPITALIZATION adds emphasis and urgency
• No special characters — sanitized by sanitizeSpeechText()

Pauses & Pacing

v3 does NOT support SSML <break> tags. Use instead:

• Ellipses (...) — adds pause and weight ("Time for a nap... zZz")
• Dashes (—) — short pause
• CAPITALIZATION — increases emphasis

Engine-Specific Behavior

Audio tags like [laughs], [sighs], [excited] are ElevenLabs v3 only. When Kokoro or Deepgram is the active engine, tags are automatically stripped before synthesis — the text "Hello!" still reaches the engine, but "[laughs]" is removed. No engine loses features because of another engine's limitations.

Kokoro supports Markdown pronunciation: [word](/IPA/) and stress: [word](+2) Deepgram Aura-2 uses natural prosody from punctuation — no markup needed.

Stability Settings (v3 exact values only)

• 0.0 (Creative): Most expressive — our default for Pompom
• 0.5 (Natural): Balanced — good for professional/mentor personalities
• 1.0 (Robust): Very stable, less responsive — good for zen personality