voiceclaw

Voice conversation interface for OpenClaw using wake word detection, streaming LLM responses, and text-to-speech. Use when a user wants to talk to their OpenClaw agent by voice, set up a voice assistant, or add speech input/output to OpenClaw. Supports configurable wake words, VOICEVOX TTS, and sentence-level streaming for low-latency responses.

Voice conversation skill for OpenClaw: wake word → STT → LLM (streaming) → TTS → playback.

Requirements

• OpenClaw running locally (gateway with chatCompletions enabled)
• Node.js 18+
• VOICEVOX running on localhost:50021 (download)
• Chrome/Edge (Web Speech API for STT)
• HTTPS for remote mic access (localhost works without HTTPS)

Quick Start

bash

# Install
git clone https://github.com/kentoku24/voiceclaw.git
cd voiceclaw
npm install

# Start (no .env needed if OpenClaw is running locally)
npm start
# → [voiceclaw] OpenClaw config loaded from ~/.openclaw/openclaw.json
# → [voiceclaw] listening on http://127.0.0.1:8788

# Open browser
open http://127.0.0.1:8788

Press 開始, say the wake word (default: アリス), then speak your command.

Configuration

All settings are optional. Set in .env or environment variables:

Variable	Default	Description
WAKE_WORDS	アリスちゃん,アリス,...	Comma-separated wake words
STT_LANG	ja-JP	Speech recognition language
OPENCLAW_MODEL	openclaw	LLM model name
VOICEVOX_URL	http://127.0.0.1:50021	VOICEVOX endpoint
VOICEVOX_SPEAKER	1	VOICEVOX speaker ID
HOST	127.0.0.1	Server bind address
PORT	8788	Server port

Gateway token is auto-detected from ~/.openclaw/openclaw.json. Override with OPENCLAW_GATEWAY_TOKEN if needed.

Architecture

text

Wake word (browser STT) → voiceclaw server → OpenClaw Gateway (streaming)
                                           → sentence-level TTS (VOICEVOX)
                                           → audio playback (Web Audio API)

See docs/architecture.md for the full sequence diagram.

API Endpoints

Method	Path	Description
GET	/health	Health check
GET	/api/config	Client-safe settings (wake words, STT lang)
POST	/api/chat-stream	Streaming LLM → sentence-level SSE
POST	/api/chat	Non-streaming LLM (fallback)
POST	/api/tts	Text → VOICEVOX → WAV audio