The model is based on a selection of late medieval German manuscripts from the fifteenth and sixteenth centuries. With a geographical focus on Austria and the Hanseatic Baltic region, both Middle Low and Early New High German texts are represented in the training material. Among the fonts used, Gothic cursive, Bastarda, and early Kurrent are most common.
The training data set is composed of approximately 77.000 words. The ground truth was compiled from several projects currently in progress at various institutions. The aim is to significantly expand and diversify this in the future to train a generic model. The current model therefore has a prototype status, with a CER of 5.60% on a validation set.
Image: WStLA 184.108.40.206.B1.6b.2 – Gewährbuch C, fol. 292r (1420)