Syllable-Guided Handwritten Text Recognition for Degraded Historical Tibetan Manuscripts

Historical Tibetan manuscripts, spanning from 8th-century Dunhuang cave scrolls to medieval monastic xylographs and administrative cursive registers, represent an essential repository of Central Asian cultural heritage. The automated transcription of these documents through Handwritten Text Recognition (HTR) is challenged by severe material deterioration, ink bleed-through, fluctuating baselines, and complex two-dimensional Brahmic syllabic stacking. In this study, we propose an end-to-end HTR architecture designed specifically for historical Tibetan document scripts. Our framework combines deformable text-line segmentation with a hybrid ResNet-Vision Transformer (ResNet-ViT) visual encoder and a joint Connectionist Temporal Classification (CTC) and autoregressive cross-attention decoder. To alleviate extreme data scarcity among infrequent vertical liga-tures, we introduce a syllable-centric synthetic data augmentation protocol. We conduct comprehensive evaluations across three authentic benchmark datasets: the Dunhuang Tibetan Document Dataset (DTDD), the Northwest Minzu University Historical Document Benchmark (NMU-THDTL), and the Tibetan Handwriting Khyug-yig Dataset (THKD). The proposed architecture achieves a Character Error Rate (CER) of 4.79% on formal woodblock dbu can (NMU-THDTL), 7.15% on archaic Dunhuang scrolls (DTDD), and 9.08% on cursive dbu med (THKD), outperforming standard baseline architectures by over 3.7% absolute CER while maintaining real-time inference throughput.