Gemini Live Avatar

Gemini Live Avatar is an open-source web application that aims to provide a conversational, real-time interface using voice, text, and animated avatars in the browser. While the user interface runs entirely in the browser, it depends on a backend server to handle WebSocket communication and interact with the Gemini Live API. Under the hood, the Gemini Live API enables seamless, low-latency interactions—allowing the 3D avatar to listen, speak, and react in real time, making conversations with AI feel more natural and engaging.

✨ Features

• ⚡ Real-time interaction powered by the Gemini Live API
• 🎤 Speech-to-Text User can interact with the avatar using voice input
• 🗣️ Text-to-Speech for the avatar's spoken responses, including lipsync and facial animations
• 💬 Text prompting with Gemini’s streaming multimodal responses
• 🧠 Avatar animation using Ready Player Me and Talking Head
• 🎥 Webcam and screen sharing capabilities for real-time context
• 📄 Multimodal chat log displaying user prompts and Gemini responses

🧠 How It Works

Gemini Live Avatar provides an interactive loop where the avatar listens, sees, responds, and reasons in real time:

1. User speaks, types, or shares screen/camera input.

2. The server receives, processes, and analyzes input streams using the Gemini Live API, which handles the request with full multimodal context—including what the avatar "sees" through shared screen or camera input.

3. The avatar responds instantly as Gemini-generated responses are received by the frontend and drive the avatar’s animation:

   • Text responses are displayed in the chat log in real time.
   • Speech responses are synthesized and played back, with the avatar lip-syncing and animating to match the spoken content.

4. Function calling is triggered dynamically when needed:

   • 🔍 Google Search grounding enhances answers with fresh external information.
   • ⚙️ Some Custom tools has been implemented, such as Turn the green(any color) lights on, turn off light, to demonstrate how function calling can be integrated into the system.

5. Screen and camera content can be referenced directly in user queries like:

   • "What’s in this slide?"
   • "Can you summarize the text on screen?"
   • "Tell me what’s in front of the camera."

This real-time loop enables expressive, grounded, and multimodal conversations with an avatar interface.

Roadmap

• End-to-end Gemini Live API integration
• Speech-to-Text & Text-to-Speech functionality
• Text input with streaming responses including multimodal content
• Webcam and screen sharing for real-time context
• Avatar animation with Ready Player Me
• Avatar animation with Mixamo
• Function calling by providing the MCP server URL
• Integrate Gemini native audio output support
• Integrate with ADK
• Add interruption support for real-time responses

Prerequisites

• Node.js v18 or later
• A Google AI Studio project with a Gemini API key
• Python 3.11+
• (Optional) Ready Player Me avatar URL

Installation

Development Setup

git clone https://github.com/haruiz/gemini-live-avatar.git
cd gemini-live-avatar
uv sync

PIP installation

pip install gemini-live-avatar

Run the App

gemini-live-avatar --google-search-grounding --workers 1 --avatar-path https://models.readyplayer.me/<AvatarID>.glb

Then open your browser at: http://localhost:8080

🧠 Using Ready Player Me

This project integrates avatars from Ready Player Me, which offers fully rigged, customizable 3D characters ideal for expressive visual representation. Facial movements—including lip sync, eye tracking, and gestures—are animated in real time using the open-source Talking Head library by Mika Suominen, and are driven by responses from the Gemini Live API. Users can personalize the experience by supplying their own Ready Player Me avatar URL.

📦 Built With

• Gemini Live API
• Vite – Modern dev environment
• Ready Player Me – Avatar creation platform
• Three.js – 3D rendering engine

🤝 Contributing

Contributions, suggestions, and pull requests are very welcome! If you'd like to contribute, please open an issue or submit a PR.