On Curating HTR Training Datasets for Romanian Language with use of Transcribathon Tool

This paper presents a workflow for HTR dataset generation in Romanian using Transcribathon’s Correct HTR feature. Leveraging citizen-science transcriptions aligned with Transkribus outputs, our case study on Jurnalul lui Dumitru Nistor reduced CER from 15.26% to 0.13%. The approach enables efficient dataset curation and supports scalable model development in low-resource languages.