🍓 Whiz — The Future of Japanese Input

Next-generation Japanese IME powered by lightweight neural networks

"Faster. Smoother. Smarter." A modern Japanese Input Method Editor featuring ConnectNet a unified neural model for superior kana-to-kanji conversion.

✨ Highlights

• ⚡ Blazing Fast — Real-time inference with a compact neural network
• 🧠 Intelligent Predictions — Context-aware word suggestions
• 🎨 Beautiful UI — Clean, minimal candidate window with Nocturne Violet theme
• 🔄 GTK3 & GTK4 Support — Works perfectly on modern Linux desktops
• ⌨️ Familiar Controls — Natural keybindings (Space, Tab, Arrow keys, etc.)
• 📦 Lightweight — Excellent accuracy with a small model footprint
• 🌍 Open Source — Fully customizable and extensible

🚀 Quick Start

Prerequisites

# On Ubuntu/Debian
sudo apt install gtk-4-dev libgtk-3-dev libpango1.0-dev

# On Fedora
sudo dnf install gtk4-devel gtk3-devel pango-devel

Installation

git clone https://github.com/yui0/whiz.git
cd whiz
make
sudo make install

Then go to Settings → Region & Language → Input Sources and add "Berry Input Method (LLM)".

🎮 How to Use

• Toggle Japanese/English: Half-width/Full-width key or Ctrl + Space
• Convert: Space or Henkan key
• Navigate Candidates: ↑ ↓ or Tab / Shift+Tab
• Move Segments: ← →
• Adjust Segment Boundaries: Shift + ← →
• Commit: Enter

Prediction mode activates automatically as you type!

🛠 Project Structure

berry-whiz/
├── im.c              # GTK3 IM module
├── im4.c             # GTK4 IM module
├── whiz.c            # Core conversion engine (ConnectNet + Lattice)
├── romaji.h          # Romaji to Kana converter
├── engine/           # Training scripts (Python)
├── model.wcnn        # Trained model
└── whiz.web.dic      # Dictionary

🧠 Technical Features

• ConnectNet: Jointly learned word and connection costs in a single model
• Lattice + Beam Search: Efficient best-path finding
• WUTF Dictionary: Fast, modern dictionary format
• Advanced Romaji Conversion: High-quality kana handling
• Prediction & Reverse Conversion: Built-in support for advanced workflows

📋 Roadmap

• Enhanced Wayland support
• Larger & faster models
• User dictionary & cloud sync
• macOS and Windows ports
• Custom theme support

🤝 Contributing

Contributions are warmly welcome! Bug reports, feature ideas, model improvements, documentation feel free to open issues or PRs.

📄 License

Mozilla Public License 2.0 — Flexible and business-friendly.

---

Built with ❤️ for a better Japanese typing experience

Project Whiz — © 2026 Yuichiro Nakada