Handwritten word recognition in low-resource languages, such as historical Norwegian, faces significant challenges due to limited annotated data and complex script variations. This paper introduces a novel zero-shot learning (ZSL) architecture for recognizing seen/unseen words in historical documents, leveraging a pre-trained ResNet-50 backbone, enhanced Temporal Pyramid Pooling (TPP), and parallel Pyramidal Histogram of Characters (PHOC) and Shapes (PHOS) branches. Unlike prior frameworks like Pho(SC)Net and ResPho(SC)Net, our approach employs independent TPP paths for each PHOC level, improving sub-word alignment. This design enhances generalization to unseen classes, achieving improvements in seen and unseen class accuracies on Norwegian and George Washington (GW) datasets. By addressing data scarcity, our method supports the preservation of endangered scripts and facilitates access to cultural heritage. The architecture’s efficiency and scalability make it a robust solution for low-resource document analysis, with potential applications in linguistic research and multilingual word recognition for historical texts.