Univerzálne modely automatickej transkripcie historických tlačí: Experimenty s platformou Transkribus na tlačiach Orbis Pictus (1798) a Adparatus ad Historiam Hungariae (1735)

This paper contributes to digital humanities research and demonstrates the use of Artificial Intelligence in the processing of written cultural heritage. It focuses on the practical application of HTR+ (Handwritten Text Recognition) tools and the Transkribus platform in the context of automatic recognition and transcription of historical prints, and examines their usability in creating universal transcription models. The aim of the paper is not to analyze or compare various tools for the transcription of handwritten texts and historical prints, but to make the evaluation of several years work with Transkribus software based on the results achieved by the authors as part of the applied research project SKRIPTOR (2020 – 2024). This research was carried out with the Transkribus Expert Client and the Transkribus App software using HTR+ and PyLaia technologies. The procedure included digitizing the Orbis Pictus (1798) textbook using the ScanTent and DocScan tools, manual text segmentation, preparation of the Ground Truth sample, and training automatic transcription models for four font types (antiqua, cursive, Fraktur, Schwabacher) and four languages (Latin, Hungarian, German, Czech). The best trained models were verified on historical prints of Orbis Pictus (1820) and Adparatus ad Historiam Hungariae (1735). Their universality was evaluated using the CER and WER indicators. The research also pointed to limiting factors in the quality of the digitized material, the layout of the text, and the appearance of specific graphemes. Testing the Transkribus Print M1 supermodel demonstrated its high accuracy and usability for historical multilingual prints.