2026/07/20 by Kuntharrgyal Khysru, Wenjie Tang, Di Jin +1
Computer Science · #Speech Recognition and Synthesis #Natural Language Processing Techniques #Topic Modeling
paper · doi:10.1093/comjnl/bxag058
Abstract Language models (LMs) are essential for automatic speech recognition (ASR). Existing research on LMs has focused primarily on major world languages such as English and Chinese, with findings difficult to adapt to other languages. Tibetan, as the official language of Tibet Autonomous Region of China, has unique key properties—grammar and morphological verbs play crucial roles in understanding—that cannot be effectively modelled by existing LMs. We propose a Tibetan language model (TLM) integrates grammatical relationships and morphological verb features. Our main contributions include: (i) analysing the importance of Tibetan grammar and morphological verbs and their impact on LMs. (ii) Proposing a suffix feature awareness method to model functional word continuation relationships in sentences. (iii) Proposing an adaptive weighting method to address rare morphological verbs, which are predominantly low-frequency words in Tibetan. Experiments show that the method considering grammatical relationships achieves 4.8% lower perplexity (PPL) than state-of-the-art method. The morphological verb weighting method achieves 7.4% lower PPL, and the method combining grammar and morphological verbs achieves 9.8% lower PPL. Our approach also outperforms existing methods on ASR tasks.