OCR that actually reads your weird old books

What it does

kraken is a Python-based OCR system that handles the full pipeline: binarization, page segmentation, layout analysis, and text recognition. It outputs to standard formats like ALTO, PageXML, and hOCR. You install via pip/pipx, download a model from its public Zenodo repository, and run commands like kraken -i image.tif image.txt binarize segment ocr.

The interesting bit

Most OCR tools assume clean, modern, left-to-right Latin text. kraken starts from the opposite assumption: historical documents, right-to-left and top-to-bottom scripts, and mixed layouts are first-class citizens. The baseline segmenter and trainable reading-order detection suggest people who actually work with manuscripts built this.

Key highlights

• Supports RTL, BiDi, and top-to-bottom scripts natively
• Fully trainable layout analysis, reading order, and character recognition
• Public model repository on Zenodo; kraken list shows available models
• Outputs ALTO, PageXML, abbyyXML, and hOCR
• Tight integration with eScriptorium for GUI-based annotation and training
• Variable recognition network architecture (unclear from docs how “variable” this is in practice)

Caveats

• No Windows support; Linux and macOS only
• PDF and multi-image TIFF/JPEG2000 require extra pip install kraken[pdf]
• Python version lock: 3.10–3.13 only, and pipx may default to something incompatible
• You must manually fetch models before doing anything useful; no built-in defaults

Verdict

Digital humanists, archivists, and anyone wrangling pre-modern or non-Latin texts should look here first. If your use case is “scan a modern English invoice,” mainstream tools are probably less friction.