Collaborative Workflows for Handwritten Text Recognition in Under-Resourced Manuscript Collections

This article addresses important questions that arise when trying to transcribe large and diverse historical manuscript collections, with a focus on under-resourced languages and scripts. Using a pilot study of challenging Tibetan manuscripts, we propose an efficient collaborative workflow that leverages existing layout recognition and HTR models and tools, including Transkribus, with iterative model training, and quantitative and qualitative error analysis. We show how this approach not only improves transcription accuracy but also provides a flexible framework adaptable to other under-resourced manuscript collections, supporting scalable text digitisation projects.