Claude Avatar

Claude Avatar

A 3D talking avatar connected to **Claude Code CLI** for real-time voice conversation — no API keys required.

Built on top of [TalkingHead](https://github.com/met4citizen/TalkingHead) by Mika Suominen (MIT License).

---

How it works

Mic → faster-whisper (GPU) → Claude Code CLI → edge-tts → Avatar lip-sync

• Voice input — MediaRecorder → ffmpeg WAV → faster-whisper large-v3-turbo on CUDA
• Claude — Persistent claude CLI session via stream-json protocol (~5s response, no startup overhead)
• Voice output — edge-tts en-US-AvaMultilingualNeural with word-level timing → TalkingHead lip-sync
• UI — Split layout: 3D avatar + chat (top) / CLI activity log (bottom)

---

Requirements

• Claude Code CLI installed and authenticated
• Python 3.10+
• Node.js 18+
• ffmpeg
• NVIDIA GPU (for faster-whisper CUDA)

pip install faster-whisper edge-tts
npm install

---

Run

node claude/server.cjs

Then open `claude/index.html` in your browser (or serve it via a local HTTP server).

---

File structure

claude/
  server.cjs           # WebSocket bridge — Claude CLI + Whisper + TTS
  index.html           # Frontend — avatar, chat, CLI log
  tts.py               # edge-tts neural voice with word boundaries
  transcribe_server.py # Persistent Whisper server (model stays in VRAM)
  transcribe.py        # One-shot Whisper fallback

---

Credits

• TalkingHead — Mika Suominen (MIT License)
• faster-whisper — GPU-accelerated transcription
• edge-tts — Microsoft neural voices
• Three.js — 3D rendering

---

License

MIT — see [LICENSE](LICENSE). Original TalkingHead library © 2023-2024 Mika Suominen.

---

Author

Built by [Agrici Daniel](https://agricidaniel.com/about) - AI Workflow Architect.

• [Blog](https://ag