Adapting TrOCR for Printed Tigrinya Text Recognition Word-Aware Loss Weighting for Cross-Script Transfer Learning

Transformer-based OCR models have achieved strong results on Latin and CJK scripts, yet their application to African syllabic writing systems remains untested. We present the first adaptation of TrOCR for printed Tigrinya text recognition using the Ge’ez script. Starting from a pre-trained TrOCR-base model, we extend the byte-level BPE tokenizer to cover 230 Ge’ez characters and introduce Word-Aware Loss Weighting, a training objective that resolves systematic word-boundary failures caused by applying Latin-centric BPE conventions to a non-Latin script. The unmodified model produces no usable output on Ge’ez text; after adaptation, the TrOCR-Printed variant achieves 0.22% Character Error Rate and 97.20% exact match accuracy, and the TrOCR-Handwritten variant achieves 0.38% CER and 96.86% exact match accuracy, on a held-out test set of 5,000 synthetic text-line images from the GLOCR dataset. A CRNN-CTC baseline trained on the same data achieves 0.12% CER and 98.20% exact match accuracy. An ablation study confirms that the weighted loss is the critical component: vocabulary extension alone yields 20% CER, while adding Word-Aware Loss Weighting reduces it by two orders of magnitude. A CRNN-CTC baseline trained on the same data slightly outperforms the TrOCR variants on printed synthetic text, providing a strong conventional benchmark for comparison. The full pipeline trains in under three hours on a single 8 GB consumer GPU. All code, model weights, and evaluation protocols are publicly released to establish the first Transformer-based benchmark for Tigrinya OCR.