PRONUNX: A PRAGMATICS-DRIVEN FRAMEWORK FOR ANALYZING PRONUNCIATION PATTERNS AND IRREGULARITIES

This study examines how Large Language Models (LLMs) process pronunciation patterns, with a focus on their capacity to handle both regularities and exceptions in spoken language. We introduce PronunX, a pragmatics-informed computational framework designed to analyze pronunciation using stress, prosodic variation, and phonological context. Unlike prior models that rely purely on phonetic rules or data-driven learning, PronunX integrates focus-sensitive structures rooted in linguistic theory with phoneme-aligned speech data. To evaluate the framework, we use the TIMIT Acoustic-Phonetic Corpus, comprising over 6,000 annotated sentences across eight major American English dialects. The speech data is processed using MFCC-based acoustic features and aligned phonetic transcriptions, and then structured for testing against four major language models: GPT-4, GPT-3.5, LLAMA-2, and BERT. Results show that while GPT-4 achieves 95.8% accuracy on regular stress patterns, performance drops markedly for irregular or context-sensitive cases. The evaluation further reveals limited prosodic sensitivity and inconsistent adaptation to dialectal variation across models. These findings underscore the need for hybrid approaches that bridge linguistic theory with adaptable learning mechanisms. Unlike previous studies that merely observe model performance gaps, this work introduces a pragmatics-driven evaluation framework—PronunX—that operationalizes stress and prosody using focus-sensitive tripartite structures. Through unsupervised testing on phoneme-aligned data, the study offers a structured, theory-informed approach to analyzing how LLMs interpret variation in speech patterns. While our corpus focuses on academic discourse, we note that Saudi Academic English may pattern differently in non-academic registers (e.g., business, healthcare, customer service); these contrasts are outlined in the Limitations and chart a clear path for future sampling.