A real-time voice AI that joins Jitsi video calls and has actual conversations.

No cloud APIs. Runs on a laptop. Fully self-hosted.

• Joins a Jitsi Meet room as a participant named "Enki"
• Listens to speech via WebRTC audio capture
• Transcribes in real-time using Whisper (~0.6s latency)
• Responds with synthesized voice using Edge TTS
• All processing happens locally

You: "Hello, can you hear me?"
Enki: "Hello! Nice to hear from you!"

You: "What's your name?"
Enki: "I am Enki, god of wisdom and water. I'm your AI assistant."

• Linux (tested on Ubuntu 24.04 / WSL2)
• Node.js 18+
• Python 3.10+
• Docker & Docker Compose
• PulseAudio
• FFmpeg

# Clone Jitsi Docker setup
git clone https://github.com/jitsi/docker-jitsi-meet.git jitsi
cd jitsi

# Configure
cp env.example .env
# Edit .env and set:
# - JVB_ADVERTISE_IPS=127.0.0.1,<your-ip>
# - ENABLE_GUESTS=1

# Start
docker compose up -d

# Clone this repo
git clone https://github.com/yaxzone/openclaw-talks-back.git
cd openclaw-talks-back

# Install Node dependencies
npm install

# Create Python venv and install Whisper
python3 -m venv venv
source venv/bin/activate
pip install faster-whisper edge-tts

pactl load-module module-null-sink sink_name=VirtualMic sink_properties=device.description=VirtualMic

node voice-bot.js

Open https://localhost:8443/Enki in your browser and start talking!

Environment variables:

The bot connects to OpenClaw's API for intelligent responses. This means you get a real AI conversation, not canned responses.

1. Enable the responses endpoint in OpenClaw config:

{
  gateway: {
    http: {
      endpoints: {
        responses: { enabled: true }
      }
    }
  }
}

2. Set your token:

export OPENCLAW_TOKEN="your-gateway-token"

3. Run the bot — it will now route speech to OpenClaw and speak the AI's response!

See ARCHITECTURE.md for detailed technical documentation including:

• System diagram
• Data flow (STT and TTS)
• Key challenges and solutions
• Performance metrics

1. ICE Connection Failures: Bot connects via localhost but JVB advertised different IP. Fixed by adding 127.0.0.1 to JVB_ADVERTISE_IPS.

2. Slow Whisper: Model reload per-chunk was too slow. Created persistent server that keeps model in memory.

3. TTS Audio Routing: Chrome's fake-media-stream only sends test patterns. Used PulseAudio virtual sinks to route TTS audio into WebRTC stream.

4. Echo Loop: Bot was transcribing its own TTS. Added isSpeaking flag to skip transcription during playback.

For persistent operation (auto-start, auto-restart on crash):

# Copy service file
sudo cp jitsi-voice-bot.service /etc/systemd/system/

# Edit the service file to set your OPENCLAW_TOKEN
sudo nano /etc/systemd/system/jitsi-voice-bot.service

# Enable and start
sudo systemctl daemon-reload
sudo systemctl enable jitsi-voice-bot
sudo systemctl start jitsi-voice-bot

# View logs
sudo journalctl -u jitsi-voice-bot -f

The bot will now:

• Auto-start on system boot
• Auto-restart if it crashes
• Always be waiting in the configured room

• Wake word detection ("Hey Enki")
• Streaming STT for lower latency
• Integration with LLM for intelligent responses ✅ Done via OpenClaw API
• Systemd service for persistence ✅ Done
• Multiple room support

MIT

Built with 🔱 — February 2026