Using AI to train Tesseract OCR that runs without AI

We are training Tesseract OCR for three languages: Ho, for which no OCR exists; Odia, for the first time, on 19th- and 20th-century fonts; and Santali. Although the entire training ran locally, building three parallel workflows and numerous specific Python scripts was possible thanks to AI.

Tesseract was trained nine years ago for Odia with modern-looking fonts. But the texts that need OCR the most are scanned books printed on a letterpress. Tesseract was also trained for Ol Chiki, the Santali-language alphabet, ten years ago. But that version of Tesseract is quite outdated, and improving a model that was trained only on synthetic renders won’t be helpful in the future. OCR does not exist for Warang Citi.

Tesseract is a very robust OCR engine. It runs locally, requiring no server or even internet connectivity when downloaded. It can serve as a foundation for building more sophisticated OCR engines. With proper training, Tesseract can be quite accurate.

My talk will focus on what went into training these three models and how a coding AI helped create language and data-specific scripts.

Key takeaways:

• How to fine-tune OCR to better read historical texts

• OCR and complex scripts

• What a typical workflow looks like for languages for poor support

• Why to even spend time and resources to train an OCR engine in the age of AI