Ка генеричком HTR моделу за српске средњовековне рукописе

Abstract

This paper presents the process of training and evaluating a general-purpose HTR (Handwritten Text Recognition) model designed for the automatic transcription of Serbian medieval manuscripts written in various forms of Cyrillic script, utilizing the Transkribus software platform. The primary practical outcome of this research is the development of the first version of a generic HTR model for Serbian medieval manuscripts, named Miroslav 1.0, in honor of the Miroslav Gospel, the most representative manuscript of the Serbian medieval tradition. The model was trained on a large and heterogeneous dataset comprising approximately 600,000 words extracted from 12th to 18th century manuscripts of various genres, all written in distinct styles of Cyrillic script (uncial, semi-uncial and cursive). The quantitative evaluation of the model’s performance indicates a character error rate (CER) ranging between 5% and 10% on out-of-sample manuscripts, which is considered highly satisfactory for historical manuscript transcription tasks. The implementation of this model significantly accelerates the transcription process of Serbian medieval manuscripts into machine-readable formats, thereby opening new avenues for corpus-based and quantitative research into the Serbian written heritage. Particularly noteworthy is the model’s extensibility: its accuracy and robustness can be further enhanced by expanding the training dataset with additional material. In addition to future improvements based on dataset augmentation, we also plan to train the model using the eScriptorium platform. This will provide researchers with an open-access alternative to Transkribus, thereby promoting broader accessibility and sustainability of digital tools in Slavic medieval manuscript studies.