Assessing advanced handwritten text recognition engines for digitizing historical documents.

Abstract

This study provides critical insights and evaluates the performance of state-of-the-art Handwritten Text Recognition (HTR) engines-PyLaia, HTR + , IDA, TrOCR-f, and Transkribus' proprietary Transformer-based "supermodel" Titan-to digitize historical documents. Using a diverse range of datasets that include different scripts, this research assesses each engine's accuracy and efficiency in handling multilingual content, complex styles, abbreviations, and historical orthography. Results indicate that, while all engines can be trained or fine-tuned to improve performance, Titan and TrOCR-f exhibit superior out-of-the-box capabilities for Latin-script documents. PyLaia, IDA, and HTR + excel in specific non-Latin scripts when specifically trained or fine-tuned. This study underscores the importance of training, fine-tuning, and integrating language models, providing critical insights for future advancements in HTR technology and its application in the digital humanities.