GitHub - sachaabot/openclaw-voice-agent: An open-source voice agent built on the PamirAI Distiller device, combining speech recognition, and text-to-speech to create a conversational AI assistant with OpenClaw you can talk to.

4 min read Original article ↗

OpenClaw Voice Agent on Distiller

Wake-word activated voice interface for OpenClaw sessions on PamirAI devices.

Flow: Listen for wake word → Capture speech → Transcribe with Whisper → Send to OpenClaw → Speak response

Requirements

  • Raspberry Pi CM5 (PamirAI Distiller)
  • Microphone and speaker connected
  • Python 3.10+ (Distiller SDK venv)
  • OpenClaw local gateway running (see below)
  • API keys: Picovoice (wake word), OpenAI (Whisper), AI provider for OpenClaw

Setup

cd /home/distiller/Projects/openclaw-voice-agent

# Install dependencies into SDK venv
source /opt/distiller-sdk/activate.sh
pip install -r requirements.txt

# Configure
cp config.example.yaml config.yaml
# Edit config.yaml with your API keys and settings

Porcupine Wake Word Setup

Get Your Access Key (Free)

  1. Visit Picovoice Console and sign up (no credit card needed)
  2. Copy your AccessKey from the dashboard
  3. Set it in config.yaml: 
    porcupine:
  access_key: "YOUR_PICOVOICE_ACCESS_KEY"

Option 1: Built-in Keywords (Easiest)

Use one of the pre-trained keywords:

  • "terminator", "porcupine", "bumblebee", "americano", "blueberry", and more

Works offline and is ready to go:

porcupine:
  keywords:
    - "jarvis"
  sensitivities:
    - 0.6

Option 2: Custom Wake Word (Better UX)

Train a custom wake word like "hey openclaw":

  1. Go to Picovoice Console
  2. Click "Create Custom Keyword"
  3. Enter your phrase (e.g., "hey openclaw")
  4. Select your platform: Raspberry Pi (32-bit ARM) (for PamirAI/CM5)
  5. Train the model (takes ~1 min)
  6. Download the .ppn file
  7. Copy it to your project directory (e.g., models/hey-openclaw_en_raspberry-pi.ppn)
  8. Update config.yaml: 
    porcupine:
  keyword_paths:
    - "/home/distiller/Projects/openclaw-voice-agent/models/hey-openclaw_en_raspberry-pi.ppn"

Fine-tune Detection

Adjust sensitivity (0.0–1.0) if needed:

  • Higher = catches quieter speech but more false positives
  • Lower = requires louder/clearer speech
  • Default (0.6) works well for most cases

TTS Providers

Provider Quality Speed Offline Setup
gtts OK Slow No None (default)
elevenlabs Excellent Fast No API key required
piper Good Fast Yes Download model

Running

# Direct
source /opt/distiller-sdk/activate.sh
python3 openclaw_voice_agent.py

# Or with custom config path
OPENCLAW_CONFIG=/path/to/config.yaml python3 openclaw_voice_agent.py

Systemd Service

Run the voice agent as a background service that survives SSH disconnects and auto-starts on boot.

Install

sudo cp openclaw-voice-agent.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable openclaw-voice-agent
sudo systemctl start openclaw-voice-agent

Manage

sudo systemctl start openclaw-voice-agent    # Start
sudo systemctl stop openclaw-voice-agent     # Stop
sudo systemctl restart openclaw-voice-agent  # Restart (e.g. after code changes)
sudo systemctl status openclaw-voice-agent   # Check if running

Logs

journalctl -u openclaw-voice-agent -f        # Tail logs live
journalctl -u openclaw-voice-agent -n 50     # Last 50 lines
journalctl -u openclaw-voice-agent --since "5 min ago"  # Recent logs

Disable

sudo systemctl stop openclaw-voice-agent
sudo systemctl disable openclaw-voice-agent  # Remove from auto-start

OpenClaw Setup

OpenClaw is a local, self-hosted Node.js gateway that connects chat platforms to AI agents. It does NOT require an OpenClaw API key; instead it uses your AI provider credentials (Anthropic, OpenAI, etc.).

Prerequisites:

  • Node.js 22+
  • AI provider API key (Anthropic, OpenAI, etc.)

Start OpenClaw:

# Install (if not already done)
npm install -g openclaw

# Configure with your AI provider key
# Edit ~/.openclaw/openclaw.json with your credentials

# Start the gateway
openclaw gateway start

# Verify it's running
openclaw status
# Should show: Gateway is running on http://localhost:18789

# List available sessions
openclaw sessions list

Documentation:

LED Feedback

The voice agent uses the PamirAI device's onboard LED to provide visual feedback during operation:

  • 🔵 BLUE - Wake word detected, listening for your voice
  • 🟢 GREEN - Processing your speech, waiting for OpenClaw response
  • 🔴 RED - Error occurred during processing
  • ⚫ OFF - Idle or task complete

The LED states help you understand what the device is doing without needing text output.

Customization

Control LED behavior in config.yaml:

led:
  enabled: true    # Enable/disable LED feedback
  index: 0         # Which LED to use (0-6 available on PamirAI)

Configuration

See config.example.yaml for all options. Key settings:

  • porcupine.access_key - Free from Picovoice Console
  • whisper.api_key - OpenAI API key
  • openclaw.base_url - Local Gateway URL (default: http://localhost:18789)
  • openclaw.agent_id - Agent to route messages to (default: main)
  • openclaw.session_id - Optional specific session ID (overrides agent_id)
  • openclaw.timeout - Max seconds to wait for agent response (default: 60)
  • tts.provider - Choose gtts, elevenlabs, or piper
  • audio.silence_threshold - Adjust if it cuts off too early or waits too long

Important: The voice agent does NOT need an OpenClaw API key. OpenClaw runs locally and uses your AI provider credentials (configured in ~/.openclaw/openclaw.json).