Automated Transcription of Gə'əz Manuscripts Using Deep Learning

This paper describes a collaborative project designed to meet the needs of communities interested in Gə'əz language texts – and other under-resourced manuscript traditions – by developing an easy-to-use open-source tool that converts images of manuscript pages into a transcription using optical character recognition (OCR). Our computational tool incorporates a custom data curation process to address the language-specific facets of Gə'əz coupled with a Convolutional Recurrent Neural Network to perform the transcription. An open-source OCR transcription tool for digitized Gə'əz manuscripts can be used by students and scholars of Ethiopian manuscripts to create a substantial and computer-searchable corpus of transcribed and digitized Gə'əz texts, opening access to vital resources for sustaining the history and living culture of Ethiopia and its people. With suitable ground-truth, our open-source OCR transcription tool can also be retrained to read other under-resourced scripts. The tool we developed can be run without a graphics processing unit (GPU), meaning that it requires much less computing power than most other modern AI systems. It can be run offline from a personal computer, or accessed via a web client and potentially in the web browser of a smartphone. The paper describes our team’s collaborative development of this first open-source tool for Gə'əz manuscript transcription that is both highly accurate and accessible to communities interested in Gə'əz books and the texts they contain.