Voice

Voice — TTS and transcription

Set up and test the agent's voice backends. Voice is OPTIONAL — off by default. See docs/voice.md for the full reference including channel-plugin precedence.

Dispatch

The agent invokes voice_speak / voice_transcribe directly when it needs to produce or consume audio during regular conversation. This skill is only for setup and diagnostics.

Setup flow

1. Call voice_status({ format: "json" }) to see current state.
2. If no TTS backend is available, suggest in order of preference:
   • sag (recommended): "Run brew install steipete/tap/sag. It's a small wrapper around ElevenLabs with good voice-prompting conventions."
   • OpenAI TTS: "Set export OPENAI_API_KEY=... in your shell rc (~/.zshrc or ~/.bashrc). Then restart the agent."
   • macOS say: "Built in on macOS. Sounds robotic but zero setup. No action needed — will be used as fallback."
3. If sag is installed but ELEVENLABS_API_KEY is missing, instruct: "Get a key from https://elevenlabs.io. Add export ELEVENLABS_API_KEY=sk_... to your shell rc. Restart the agent."
4. If the user also wants STT, same logic: recommend whisper-cli (brew install whisper-cpp, offline, free) or OpenAI Whisper API (same OPENAI_API_KEY).
5. Enable voice in config:
   • Tell the user: "I'll set voice.enabled: true in your config. Run agent_config(action='set', key='voice.enabled', value='true') or edit agent-config.json directly."
   • (You MAY call agent_config for them after they confirm.)
6. If the sag skill is in an OpenClaw workspace (~/.openclaw/workspace*/skills/sag/), offer: "I see you have the sag skill in an OpenClaw workspace. Want me to install it into this agent? Run /agent:skill install <that path>."
7. Mention WhatsApp precedence if the WhatsApp plugin is configured:
   • If voice_status reports whatsapp.audioEnabled: true — "Your WhatsApp plugin already transcribes voice notes locally. For inbound WhatsApp audio you don't need voice_transcribe. Setting this up is for WebChat uploads, iMessage audio, outbound voice notes, etc."
   • If false — "Your WhatsApp plugin doesn't transcribe by default. Either turn that on with /whatsapp:configure audio (local Whisper, free), or use our voice_transcribe per message."

Test flow

1. Call voice_status to confirm voice is enabled AND a backend is available. If not, redirect to setup.
2. Call voice_speak({ text: "<greeting in user's language>" }).
3. Print the result: "Generated at /tmp/... using backend X. Play it or attach it in a messaging channel."
4. If the user is on a messaging channel that supports media, use the channel's own reply/send tool with the path (e.g. MEDIA:/tmp/... or a dedicated send_media tool).

Secrets handling

Never put API keys in agent-config.json. They are SECRETS and the config file may end up in a git repo. Use environment variables:

export ELEVENLABS_API_KEY=sk_...
export OPENAI_API_KEY=sk-...

Add them to ~/.zshrc / ~/.bashrc / ~/.config/fish/config.fish for persistence.

Non-secret settings (default backend, voice ID, output dir) go in agent-config.json.

Channel-plugin precedence (IMPORTANT)

A channel plugin that already transcribes audio (like the WhatsApp plugin with audio on) is authoritative for THAT channel. Do not call voice_transcribe on an audio file that arrived through such a plugin — you'd just be re-doing work the plugin already did, and you'd end up with two different transcriptions.

voice_transcribe is for:

• Files from channels without built-in transcription (WebChat uploads, iMessage attachments)
• Standalone files (the user dropped an .m4a in the workspace)
• Cases where the channel plugin's transcription is disabled

Response style

• /agent:voice status on CLI: full card.
• On messaging channels: compact — single line per backend, skip the detailed reasons.
• /agent:voice setup: step-by-step, one instruction at a time. Don't dump all 7 steps — guide the user through what's missing for their situation.

Never

• Don't write API keys to agent-config.json — they're env-only.
• Don't auto-run brew install — always let the user do it.
• Don't transcribe WhatsApp voice notes with voice_transcribe if the plugin has audio-on.

References

• docs/voice.md — full doc (backends, precedence, secrets, troubleshooting)
• lib/voice.ts — routing + backends
• skills/skill-manager/SKILL.md — for installing the sag skill from OpenClaw path