Enable Voice

By the end of this page, you'll be able to speak to your agent and hear it respond out loud.

Time: ~30 minutes

What Talk Mode is

Talk Mode is OpenClaw's voice interface. It works like this:

1. You press a wake key (or say a wake word)
2. You speak — OpenClaw transcribes your speech using Whisper (STT)
3. Your agent processes the message and generates a response
4. The response is spoken aloud via ElevenLabs or system TTS

It runs locally on your machine. No cloud pipeline — audio goes from mic → Whisper → your AI provider → TTS → speakers.

Prerequisites

• A microphone (built-in laptop mic works)
• An ElevenLabs API key (free tier is sufficient) — or use macOS system TTS as a free alternative
• Python 3.10+ (for Whisper — the installer handles this)

Step 1 — Enable Talk Mode

Open ~/.openclaw/openclaw.json:

{
  "talk_mode": {
    "enabled": true,
    "stt": {
      "provider": "whisper",
      "model": "base"
    },
    "tts": {
      "provider": "elevenlabs",
      "apiKey": "YOUR_ELEVENLABS_API_KEY",
      "voiceId": "YOUR_VOICE_ID"
    },
    "wakeKey": "ctrl+space"
  }
}

STT models: tiny (fastest, less accurate), base (good balance), small (slower, more accurate). Start with base.

TTS provider options:

• elevenlabs — high quality, requires API key, free tier = 10,000 chars/month
• system — uses macOS say command, free but robotic

Step 2 — Get an ElevenLabs voice ID

1. Create a free account at elevenlabs.io
2. Go to Voices → browse the default voices
3. Click a voice → copy the Voice ID from the URL

Or use the ElevenLabs API to list available voices:

curl -H "xi-api-key: YOUR_KEY" https://api.elevenlabs.io/v1/voices | jq '.voices[] | {name: .name, id: .voice_id}'

Step 3 — Install Whisper dependencies

openclaw plugins install voice

This installs the voice plugin, including the Python Whisper package, and downloads the selected model. For base, expect ~150MB download.

Step 4 — Restart and test

openclaw gateway restart

Then start the voice plugin:

openclaw plugins run voice

Press ctrl+space (or your configured wake key), speak a sentence, release. Wait for the response. You should hear your agent reply out loud.

Use a wake word instead

If you want always-on listening instead of a hotkey:

{
  "talk_mode": {
    "wakeWord": "hey claw",
    "wakeKey": null
  }
}

Wake word detection uses a lightweight local model — it doesn't send audio to any server until after the wake word is detected.

Troubleshooting