Digitizing Nepal's Written Heritage: A Comprehensive HTR Pipeline for Old Nepali Manuscripts

This paper presents the first end-to-end pipeline for Handwritten Text Recognition (HTR) for Old Nepali, a historically significant but lowresource language. We adopt a line-level transcription approach and systematically explore encoder-decoder architectures and data-centric techniques to improve recognition accuracy. Our best model achieves a Character Error Rate (CER) of 4.9%. In addition, we implement and evaluate decoding strategies and analyze tokenlevel confusions to better understand model behavior and error patterns. Although the evaluation dataset is confidential, we release our training code, model configurations, and evaluation scripts to support further research on HTR for low-resource historical scripts. However, the historical manuscripts considered in this study present several challenges, including diverse handwriting styles, degraded document quality, intricate conjunct forms, and limited annotated data (Nockels et al., 2024) . These constraints necessitate approaches that leverage transfer learning, data augmentation, and carefully designed model architectures (Garces Arias et al., 2023). In this paper, we address these challenges by conducting a comprehensive exploration of modern HTR techniques for Old Nepali manuscripts. We investigate transfer learning strategies to enable effective learning in low-resource settings, implementing a three-stage approach that adapts models from high-resource settings to target-domain data.