ra-atr

Collaborative handwritten text recognition for historical documents from Sweden's National Archives (Riksarkivet).

Live demo: https://huggingface.co/spaces/carpelan/lejonet

Layer	Technology
Frontend	Svelte 5, Tailwind CSS 4, ONNX Runtime Web
Docs	Zensical
CI/CD	Dagger, Docker, GitHub Actions

Quick Start

make setup   # Install dependencies
make dev     # Start frontend dev server at http://localhost:5173

Documentation

• Docs site — architecture, configuration, getting started

Features

• In-browser HTR — ONNX Runtime Web (WASM) with multi-threaded inference
• GPU acceleration — Optional GPU server for faster inference
• Export — ALTO XML, PAGE XML, JSON, plain text
• i18n — Swedish and English

HTR Pipeline

Layout Detection (RTMDet) → Line Detection (YOLO) → Transcription (TrOCR)

Requirements

• Node.js 22+ — frontend runtime

License

Apache 2.0