OpenClaw Voice
Open-source browser-based voice interface for AI assistants.
Talk to your AI like you talk to Alexa — but self-hosted, private, and connected to your own agent.
🌐 Website: openclawvoice.com
Features
| Feature | Description |
|---|---|
| 🎤 Local STT | Whisper runs locally via faster-whisper. Your voice never leaves your machine. |
| 🔊 Streaming TTS | ElevenLabs with sentence-by-sentence streaming. Hear responses while they generate. |
| 🎯 Voice Activity Detection | Silero VAD filters background noise. Works in noisy environments. |
| 🧹 Smart Text Cleaning | Strips markdown, hashtags, URLs before TTS. No more "hash hash". |
| 🔌 Any AI Backend | OpenAI, Claude, or full OpenClaw agent with memory and tools. |
| 🌐 Browser-Based | No app install. Works on desktop and mobile. |
| 🚗 Continuous Mode | Hands-free conversation. Auto-listens after each response. |
Quick Start
# Clone
git clone https://github.com/Purple-Horizons/openclaw-voice.git
cd openclaw-voice
# Install
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# Optional: Install VAD support (better noise handling)
pip install torch torchaudio
# Configure (create .env from example)
cp .env.example .env
# Edit .env with your API keys
# Run
PYTHONPATH=. python -m src.server.main
# Open http://localhost:8765
For OpenClaw Agents (Auto-Install)
One-liner for AI agents to self-install:
cd /tmp && \
git clone https://github.com/Purple-Horizons/openclaw-voice.git && \
cd openclaw-voice && \
python3 -m venv .venv && source .venv/bin/activate && \
pip install -r requirements.txt torch torchaudio && \
PYTHONPATH=. ELEVENLABS_API_KEY="$ELEVENLABS_API_KEY" OPENAI_API_KEY="$OPENAI_API_KEY" \
nohup python -m src.server.main > /tmp/voice-server.log 2>&1 &
Configuration
Environment Variables
| Variable | Required | Default | Description |
|---|---|---|---|
ELEVENLABS_API_KEY | Yes* | — | ElevenLabs API key for TTS |
OPENAI_API_KEY | Yes* | — | OpenAI API key (if not using gateway) |
OPENCLAW_GATEWAY_URL | No | — | OpenClaw gateway URL for full agent |
OPENCLAW_GATEWAY_TOKEN | No | — | Gateway auth token |
OPENCLAW_PORT | No | 8765 | Server port |
OPENCLAW_STT_MODEL | No | base | Whisper model size |
OPENCLAW_STT_DEVICE | No | auto | Device: auto, cpu, cuda, mps |
OPENCLAW_REQUIRE_AUTH | No | false | Require API keys for clients |
*One of OPENAI_API_KEY or OPENCLAW_GATEWAY_URL required.
Whisper Model Sizes
| Model | Speed | Quality | VRAM | Best For |
|---|---|---|---|---|
tiny | Fastest | Fair | ~400MB | Quick testing |
base | Fast | Good | ~1GB | Default. Good balance. |
small | Medium | Better | ~2GB | Clearer transcription |
medium | Slower | Great | ~5GB | Accuracy priority |
large-v3-turbo | Slow | Best | ~6GB | Maximum accuracy |
TTS Options
| Backend | Type | Quality | Latency | Notes |
|---|---|---|---|---|
| ElevenLabs | Cloud | Excellent | ~500ms | Default. Streaming supported. |
| Chatterbox | Local | Very Good | ~1s | MIT license, voice cloning |
| XTTS-v2 | Local | Excellent | ~1s | Voice cloning supported |
| Mock | Local | None | 0ms | For testing (silence) |
ElevenLabs uses eleven_turbo_v2_5 for fastest response.
OpenClaw Gateway Integration
Connect to your full OpenClaw agent (same memory, tools, and persona as text chat):
# .env
OPENCLAW_GATEWAY_URL=http://localhost:18789
OPENCLAW_GATEWAY_TOKEN=your-token
ELEVENLABS_API_KEY=your-key
Add to your openclaw.json:
{
"gateway": {
"http": {
"endpoints": {
"chatCompletions": { "enabled": true }
}
}
},
"agents": {
"list": [
{
"id": "voice",
"workspace": "/path/to/workspace",
"model": "anthropic/claude-sonnet-4-5"
}
]
}
}
Architecture
┌─────────────┐ WebSocket ┌─────────────────────────────────────┐
│ Browser │◄────────────►│ Voice Server │
│ (mic/spk) │ │ │
└─────────────┘ │ ┌─────────┐ ┌─────┐ ┌─────────┐ │
│ │ Whisper │→│ AI │→│ElevenLabs│ │
│ │ (STT) │ │ │ │ (TTS) │ │
│ └─────────┘ └─────┘ └─────────┘ │
│ ↑ │ │
│ [VAD] [streaming] │
└─────────────────────────────────────┘
Streaming Flow:
- User speaks → Whisper transcribes locally
- AI responds (streamed) → buffer sentences
- First sentence complete → TTS starts immediately
- Audio streams to browser while AI continues
- Result: ~50% faster perceived response
HTTPS for Mobile
Mobile browsers require HTTPS for microphone access. Options:
Tailscale Funnel (easiest):
tailscale funnel 8765
# Access via https://your-machine.tailnet-name.ts.net
nginx + Let's Encrypt:
server {
listen 443 ssl;
server_name voice.yourdomain.com;
location / {
proxy_pass http://127.0.0.1:8765;
proxy_http_version 1.1;
proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection "upgrade";
}
}
API
WebSocket Protocol
Connect to ws://localhost:8765/ws:
// Start recording
{ "type": "start_listening" }
// Send audio (base64 PCM float32, 16kHz)
{ "type": "audio", "data": "base64..." }
// Stop recording
{ "type": "stop_listening" }
// Receive events:
{ "type": "transcript", "text": "...", "final": true }
{ "type": "response_chunk", "text": "..." } // Streaming text
{ "type": "audio_chunk", "data": "...", "sample_rate": 24000 } // Streaming audio
{ "type": "response_complete", "text": "..." } // Full response
{ "type": "vad_status", "speech_detected": true } // VAD feedback
Roadmap
- WebSocket voice gateway
- Whisper STT (local)
- ElevenLabs TTS
- Streaming TTS (sentence-by-sentence)
- Voice Activity Detection (Silero)
- Text cleaning (markdown/hashtags/URLs)
- Continuous conversation mode
- OpenClaw gateway integration
- WebRTC for lower latency
- Voice cloning UI
- Docker support
License
MIT License — see LICENSE.
Credits
- faster-whisper — Local STT
- ElevenLabs — Text-to-Speech
- Silero VAD — Voice Activity Detection
- Built for OpenClaw
Made with 🦞 by Purple Horizons