Japan's National Diet Library Ships NDLOCR-Lite to Read Handwritten Japanese Without a GPU
Japan's National Diet Library ships a CPU-only Japanese OCR that now handles handwritten and vertical text, with open weights and CC BY 4.0.
Takeaways
- NDLOCR-Lite v1.2.1 adds stronger handwritten Japanese recognition, trained on political history archives.
- Runs CPU-only on Windows, macOS (Apple Silicon) and Ubuntu, no GPU required.
- Reports 0.268 mean CER on 1,065 handwritten images from the JaWildText benchmark.
- Pipeline uses DEIMv2 for layout, PARSeq for line recognition, exported to ONNX.
- Ships both a desktop GUI with screen-capture OCR and a Python CLI with batch mode.
- Released under CC BY 4.0 on GitHub, commercial use allowed with attribution.
The National Diet Library of Japan has pushed a new version of NDLOCR-Lite, a lightweight OCR app that runs on a regular laptop and now handles handwritten Japanese much better than before. Version 1.2.1 adds training on political history archives, including handwritten constitutional documents and occupation-era typewriter records, and the team has published concrete character-error-rate numbers on an external handwriting benchmark.
The pitch is simple: Japanese OCR without a GPU, without a cloud API, and without a restrictive license. The previous NDLOCR pipeline required an NVIDIA card, which shut out most laptop users and anyone running Mac or ARM hardware. This lite build strips that requirement while keeping the National Diet Library's accumulated training data and pipeline design.
What actually shipped
NDLOCR-Lite is distributed both as a desktop app (Windows 11, macOS on Apple Silicon, Ubuntu 22.04) and as a Python CLI. The desktop build includes a screen-capture mode that lets you OCR a region of your display directly, without exporting an image first. The tool supports Japanese printed text, handwriting, and vertical writing on a CPU-only machine.