Predicting Phonetic Changes Using Transformer-Based Phoneme Models
Abstract
Predicting phonetic changes is a crucial aspect of historical linguistics, aiding in the reconstruction of language evolution. This study presents a Transformer-Based Phoneme Model (HL-TPM) to analyze and predict phonetic shifts over time with improved accuracy. Existing methods often rely on rule-based or statistical approaches, which struggle with capturing complex, long-range dependencies in phonetic transformations. These limitations lead to reduced predictive performance and inefficiencies in modeling diachronic phonological changes. To address these challenges, HL-TPM leverages Transformer-Based deep learning architectures, which excel at sequence-to-sequence learning. The model is trained on historical phonetic data, utilizing self-attention mechanisms to identify patterns and predict pronunciation shifts in different linguistic lineages. Key techniques include contextual phoneme embedding, phonetic sequence alignment, and cross-temporal feature extraction. The proposed method enhances the study of language evolution by providing a robust, data-driven approach to phoneme prediction. It facilitates the reconstruction of proto-languages, improves phonetic annotation in historical texts, and aids linguistic researchers in tracing phonological shifts. Experimental results show that HL-TPM outperforms traditional models in accuracy and generalization, effectively predicting phonetic changes across languages. These findings demonstrate the potential of Transformer models in advancing computational historical linguistics.