A Speech-to-Knowledge Framework for Archaic Thai Language Understanding and Knowledge Graph Construction

Digitizing oral cultural heritage necessitates computational methods capable of bridging the acoustic, orthographic, and semantic gaps between historical and contemporary language forms. This study proposes a cascaded Automatic Speech Recognition–Neural Machine Translation–Knowledge Graph (ASR–NMT–KG) framework designed to transcribe archaic Thai recitations, translate them into modern Thai, and subsequently convert their semantic content into structured knowledge. The framework was evaluated utilizing the Thai Ancient Medical Speech Corpus (TAMSC), which comprises 14000 original and augmented utterances from ten certified Thai Traditional Medicine practitioners. The domain-adapted ASR model achieved a word error rate (WER) of 8.3%, a character error rate (CER) of 4.8%, and an Archaic Orthography Error Rate (AOER) of 6.2%. The mT5-Large translation model obtained a BLEU score of 51.7, while the Knowledge Graph construction module achieved an F1-score of 0.89. Furthermore, the WER exhibited a strong negative association with translation quality (Pearson’s r = −0.87, p < 0.001), providing empirical evidence of cross-stage performance dependency within the cascaded pipeline. These results indicate that the cascaded architecture offers an accurate, modular, and interpretable approach for preserving and semantically exploring low-resource historical language materials.