Say It Out Loud.
Your Agent Listens.

Voice Features by Platform

Voice Mode capabilities differ slightly across platforms. Here's exactly what works where.

🔊 Set Up ElevenLabs Text-to-Speech

ElevenLabs gives your agent a natural, human-sounding voice. It streams audio incrementally so replies start playing within about a second of generation.

{
  "voice": {
    "tts": {
      "provider":    "elevenlabs",
      "apiKey":      "sk_xxxxxxxxxxxxxxxx",
      "voiceId":     "21m00Tcm4TlvDq8ikWAM",  // Rachel
      "modelId":     "eleven_turbo_v2_5",        // fast + cheap
      "outputFormat":"mp3_44100_128",
      "streaming":   true                       // incremental playback
    }
  }
}

eleven_turbo_v2_5    // ← Recommended. Fast, cheap, great quality
eleven_multilingual_v2 // Best quality, supports 29 languages, slower
eleven_monolingual_v1  // Older, English only — avoid unless on free tier

Complete Voice Config

The full voice block in openclaw.json — every option in one place with comments.

{
  "voice": {

    // ── Speech-to-Text ──────────────────────────────────
    "stt": {
      "provider": "whisper",       // "whisper" | "native" | "deepgram"
      "model":    "whisper-1",
      "apiKey":   "sk-proj-xxxx",
      "language": "en"              // omit for auto-detect
    },

    // ── Text-to-Speech ──────────────────────────────────
    "tts": {
      "provider":     "elevenlabs",   // "elevenlabs" | "openai" | "native"
      "apiKey":       "sk_xxxx",
      "voiceId":      "21m00Tcm4TlvDq8ikWAM",
      "modelId":      "eleven_turbo_v2_5",
      "outputFormat": "mp3_44100_128",
      "streaming":    true,          // start playing before full text generated
      "speed":        1.0            // 0.7–1.2 playback speed
    },

    // ── Talk Mode ───────────────────────────────────────
    "talkMode": {
      "enabled":           true,
      "silenceTimeoutMs":  1500, // pause length before sending (ms)
      "interruptOnSpeech": true, // speaking stops current TTS playback
      "pushToTalk":        false // true = hold key/button instead of VAD
    },

    // ── Wake Word ───────────────────────────────────────
    "wakeWord": {
      "enabled":            true,
      "phrases":            ["hey openclaw", "hey claw"],
      "confirmationSound":  true,
      "listenDurationMs":   8000 // how long to listen after wake word (ms)
    },

    // ── Quiet Hours ─────────────────────────────────────
    "quietHours": {
      "enabled": true,
      "start":   "23:00",  // local time
      "end":     "07:00"   // TTS is silenced during this window
    }

  }
}

Get the Best Out of Voice Mode

• 🎙️
  Use a headset or AirPods for Talk Mode Built-in laptop speakers cause echo and feedback loops where the mic picks up TTS output. Any earbuds eliminate this entirely — Talk Mode works much better with audio isolated to your ears.
• 🔇
  Set quiet hours so TTS doesn't play at 3am If you use HEARTBEAT.md for automated tasks, the agent may reply via TTS even when you're asleep. Set quietHours in your config to silence audio between 23:00 and 07:00.
• ⏸️
  Enable interruptOnSpeech for natural conversation With "interruptOnSpeech": true, speaking while the agent is talking immediately stops playback and starts listening. This makes conversations feel natural instead of waiting for the agent to finish.
• 🌐
  Set the language for faster Whisper transcription If you always speak English, set "language": "en" in your STT config. Whisper skips language detection and transcribes faster. Supports 50+ languages — use your ISO code (e.g. "fr", "de", "ja").
• 🎭
  Pick a distinct wake phrase to avoid false triggers "Hey OpenClaw" is reliable, but if you often say "hey" in conversation, add something more unique like "ok lobster" or your agent's custom name. Fewer false triggers means less frustration.
• 🔋
  Disable wake word on battery power Always-on wake word detection uses a small but constant amount of CPU. If you're on a laptop away from power, set a keyboard shortcut for push-to-talk instead to save battery.

Related Topics