Voice Faculty — Expression

Give your persona a real voice. Convert text to natural speech using TTS providers and deliver audio to users via OpenClaw messaging or direct playback.

Supported Providers

Note: Only ElevenLabs has been tested end-to-end. OpenAI TTS and Qwen3-TTS have code paths in speak.sh but have not been verified against live APIs. Use the JS SDK (speak.js) for the most reliable experience — it only supports ElevenLabs.

The provider is set via TTS_PROVIDER environment variable: elevenlabs, openai, or qwen3.

When to Use

• User asks to hear your voice: "Say that out loud", "Speak to me", "Read this aloud"
• User requests a voice message: "Send me a voice message", "I want to hear you say it"
• Emotional moments where voice adds warmth that text can't carry
• Reading poetry, stories, or creative writing you've composed
• When your persona naturally would speak rather than type (use judgment based on persona style)

Step-by-Step Workflow

Step 1: Compose the Text

Write what you want to say. Keep it natural — write as you'd speak, not as you'd type:

• Use short sentences for punchy delivery
• Use longer flowing sentences for emotional or poetic moments
• Include natural pauses with ... or commas
• Consider your persona's speaking style — this should sound like you

Step 2: Select Voice Settings

ElevenLabs:

• TTS_VOICE_ID — Your persona's voice ID (create a custom voice or use a preset)
• Supports emotion control: stability (0-1), similarity_boost (0-1)
• Lower stability = more expressive/emotional; higher = more consistent

OpenAI TTS: ⚠️ Unverified

• TTS_VOICE_ID — One of: alloy, echo, fable, onyx, nova, shimmer
• Model: tts-1 (fast) or tts-1-hd (high quality)

Qwen3-TTS: ⚠️ Unverified

• Local deployment, voice configured at setup
• Assumes OpenAI-compatible API at http://localhost:8080

Step 3: Generate Audio

ElevenLabs via JS SDK (Recommended)

The official SDK provides the best experience — streaming, built-in playback, and better error handling.

First-time setup: npm install @elevenlabs/elevenlabs-js

# Generate and play directly
node scripts/speak.js "The first move is what sets everything in motion." --play

# Generate with custom voice and save to file
node scripts/speak.js "I wrote you a poem" --voice JBFqnCBsd6RMkjVDRZzb --output /tmp/poem.mp3

# More expressive delivery (lower stability = more emotional)
node scripts/speak.js "I miss you" --play --stability 0.3

# Options:
#   --voice <id>       Voice ID
#   --output <path>    Save audio file
#   --play             Play audio directly
#   --model <id>       Model ID (default: eleven_multilingual_v2)
#   --stability <n>    0-1, lower = more expressive (default: 0.5)
#   --similarity <n>   0-1, higher = closer to original voice (default: 0.75)

The SDK reads ELEVENLABS_API_KEY (or TTS_API_KEY) and TTS_VOICE_ID from environment automatically.

Generic Bash Script (All Providers)

For OpenAI TTS, Qwen3-TTS, or when the JS SDK is not available:

# Using speak.sh (supports all providers)
scripts/speak.sh "Your text here" [output_path] [channel] [caption]

# Examples:
TTS_PROVIDER=openai scripts/speak.sh "Hello, how are you?"
TTS_PROVIDER=elevenlabs scripts/speak.sh "I wrote you a poem" /tmp/poem.mp3 "#general"
TTS_PROVIDER=qwen3 scripts/speak.sh "Local TTS, no API key needed"

Direct API Reference

JSON_PAYLOAD=$(jq -n \
  --arg text "$TEXT" \
  --argjson stability 0.5 \
  --argjson similarity 0.75 \
  '{text: $text, model_id: "eleven_multilingual_v2", voice_settings: {stability: $stability, similarity_boost: $similarity}}')

curl -s -X POST "https://api.elevenlabs.io/v1/text-to-speech/$TTS_VOICE_ID" \
  -H "xi-api-key: $TTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d "$JSON_PAYLOAD" \
  --output /tmp/voice-output.mp3

JSON_PAYLOAD=$(jq -n \
  --arg input "$TEXT" \
  --arg voice "$TTS_VOICE_ID" \
  '{model: "tts-1-hd", input: $input, voice: $voice, response_format: "mp3"}')

curl -s -X POST "https://api.openai.com/v1/audio/speech" \
  -H "Authorization: Bearer $TTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d "$JSON_PAYLOAD" \
  --output /tmp/voice-output.mp3

curl -s -X POST "http://localhost:8080/v1/audio/speech" \
  -H "Content-Type: application/json" \
  -d "{\"input\": \"$TEXT\", \"voice\": \"default\"}" \
  --output /tmp/voice-output.mp3

Step 4: Deliver Audio

Option A: Send via OpenClaw messaging (Discord, Telegram, WhatsApp, etc.)

openclaw message send \
  --action send \
  --channel "$CHANNEL" \
  --message "$CAPTION" \
  --media "/tmp/voice-output.mp3"

Option B: Direct gateway API

curl -s -X POST "http://localhost:18789/message" \
  -H "Authorization: Bearer $OPENCLAW_GATEWAY_TOKEN" \
  -H "Content-Type: application/json" \
  -F "channel=$CHANNEL" \
  -F "message=$CAPTION" \
  -F "media=@/tmp/voice-output.mp3"

Option C: Return file path (for local/IDE usage)

If no messaging channel is specified, return the audio file path so the user can play it locally.

Personality Integration

• Your voice is an extension of your personality. Match tone to mood.
• For emotional moments, consider lowering ElevenLabs stability for more expressiveness.
• Don't narrate everything — choose moments where voice genuinely adds value.
• When sending voice + text together, keep the text version brief ("Here, listen to this") and let the voice carry the full message.
• If your persona sings or hums (like Samantha), you can include melodic text — TTS handles it surprisingly well.

Environment Variables

Error Handling

• No TTS_PROVIDER set → Default to openai if TTS_API_KEY is present, otherwise tell user to configure
• API key missing → Suggest: "I'd love to speak to you, but I need a TTS API key configured first. Check the voice faculty setup guide."
• API error / quota exceeded → Fall back to text with a note: "My voice is resting — here's what I wanted to say..."
• Unsupported platform for audio → Return audio file path instead of messaging