🧠 shefixu — Your AI Girlfriend Lives in Your Browser Now

She talks. She listens. She lip-syncs. She blinks. She follows your eyes. And she runs 100% on your machine.

Built from light and code — an open-source 3D companion that makes the past feel like dial-up.

👇 See it in action before you read another word

From HN headline to published LinkedIn post — all from your phone, in seconds.

---

⚡ What you're looking at

Aria is a fully animated 3D anime avatar that:

• 🗣️ Speaks back to you — in a real Microsoft neural voice, generated offline, on your GPU
• 👂 Hears you — press mic, talk, she listens and responds
• 👄 Lip-syncs in real time — every syllable, every frame, driven by raw audio frequency analysis
• 👀 Watches you — her head tracks your mouse like she's actually paying attention
• 😮‍💨 Breathes, sways, blinks — fully procedural idle animation. No animation clips. Pure math.
• 🧠 Thinks with a frontier LLM — DeepSeek, Gemini, Claude, GPT-4o — your choice, one line of code
• 💸 Costs you nothing — local TTS, free-tier LLM, zero subscriptions

This is not a chatbot with a face pasted on it. Every system — voice, motion, lip-sync, attention — is wired together in a real-time 60fps loop.

---

🔥 Why people are obsessing over this

Most "AI companion" projects are:

• A chat window with an avatar image
• Cloud-dependent, expensive, or both
• Impossible to customise without a PhD

shefixu is none of that.

It's a fully local, fully open, fully hackable 3D AI companion you can run on your laptop right now. Change her personality in one file. Swap her voice in two lines. Give her a different face by dropping a new VRM file into /public. Make her speak Japanese. Make her a therapist. Make her a pirate. She doesn't care — you're in control.

---

🏗️ How it actually works

You speak →  Web Speech API (browser-native, zero setup)
                  ↓
         OpenRouter LLM API (DeepSeek / Claude / GPT-4o)
                  ↓
         tts_server.py  →  edge-tts  →  Microsoft neural voice (offline)
                  ↓
         Web Audio API frequency analyser
                  ↓
         VRM mouth morph targets  →  real-time lip-sync  →  60fps

Every frame, the animation loop reads audio frequency data and drives Aria's mouth, spine, arms, and head simultaneously. It's the same technique AAA games use — just in a browser tab.

---

🛠️ Full tech stack

What	How
3D avatar rendering	React Three Fiber + @pixiv/three-vrm
Animations	100% procedural — useFrame() loop, no clips
LLM brain	OpenRouter — swap any model in 1 line
Voice (TTS)	edge-tts — Microsoft en-US-AnaNeural, runs fully offline
Mic input (STT)	Web Speech API — browser-native, nothing to install
Lip-sync	Web Audio API frequency → VRM morph targets, every frame
UI & transitions	React 19 + Framer Motion
Bundler	Vite

---

🚀 Get her running in under 5 minutes

What you need

• Node 18+
• Python 3.11+
• A free OpenRouter account (takes 30 seconds to sign up)

Step 1 — Clone & install

git clone https://github.com/Lumacodes/shefixu-ai-girlfriend.git
cd shefixu-ai-girlfriend
npm install
pip install edge-tts fastapi uvicorn

Step 2 — Paste your API key

cp .env.example .env

VITE_OPENROUTER_API_KEY=sk-or-v1-your-key-here

Step 3 — Drop in a VRM model

Grab any free VRM 0.x model from VRoid Hub, place it in public/, rename it model.glb. Done.

Step 4 — Start it

# Terminal 1
python3.11 tts_server.py

# Terminal 2
npm run dev

Open http://localhost:5173 and say hello.

---

🎛️ Make her yours

Give her a different brain

src/hooks/useAI.ts — swap in any OpenRouter model:

model: 'deepseek/deepseek-v4-flash',         // 🔥 default — blazing fast & free
model: 'google/gemini-2.0-flash-lite:free',  // 🆓 another free option
model: 'anthropic/claude-3.5-sonnet',        // 🧠 highest quality conversations
model: 'openai/gpt-4o',                      // 🏆 the classic

Give her a different voice

src/hooks/useSpeech.ts:

const VOICE = 'en-US-AnaNeural';  // any edge-tts voice — 300+ options
const PITCH = '+10Hz';
const RATE  = '+12%';

edge-tts --list-voices  # see all 300+ voices across 70+ languages

Change her personality

Edit the system prompt in src/hooks/useAI.ts. One paragraph of text and she becomes a completely different character — therapist, tutor, roleplay partner, language practice buddy, whatever you want.

---

📁 Codebase is tiny and clean

src/
├── components/
│   ├── ChatUI.tsx       ← chat panel, mic button, interrupt handling
│   ├── Experience.tsx   ← Three.js scene — lighting, camera, floor
│   └── Model.tsx        ← VRM loader + every animation system
├── hooks/
│   ├── useAI.ts         ← LLM client (OpenRouter)
│   └── useSpeech.ts     ← TTS client + Web Audio analyser
tts_server.py            ← 50-line FastAPI wrapper for edge-tts

~500 lines of TypeScript. You can read the whole thing in 20 minutes and understand every moving part.

---

⚠️ Known limitations (being worked on)

• TTS server must be running locally — no cloud fallback yet
• VRM 1.0 support is untested — use VRM 0.x models
• Lip-sync is frequency-based, not phoneme-based (sounds great, not perfect)
• No persistent memory between sessions yet

---

🤝 Want to make it better?

The most wanted contributions right now:

• Phoneme-based lip-sync — map viseme morphs to actual phonemes from TTS
• VRM 1.0 support
• Persistent memory — give her a JSON memory of past conversations
• Emotion system — drive face morph targets based on LLM sentiment

Open a PR. Every contribution gets acknowledged in the readme.

---

📄 License

MIT — free for personal projects, side projects, commercial projects, whatever. Go wild.

---

Built by Lumacodes

⭐ Star this if it blew your mind — it's the best way to help others find it

If this saved you weeks of research or gave you an idea for something cool — a coffee means a lot 🙏