Simulation of Teacher-Learner Interaction in English Language Pronunciation Learning

Second language (L2) pronunciation learning is a complex process involving dynamic interactions between teachers and learners. Despite its importance, computational modelling of these interactions remains limited. This PhD thesis proposed an integrated computational model that simulates both teacher and learner roles in L2 English pronunciation learning for Mandarin speakers. Validation of the model uses a purpose-built corpus, the Teacher and L2 Learner Interaction (TL2I) corpus, which includes repetition-based pronunciation improvement data and pronunciation perception evaluation data. Statistical analysis shows moderate to strong positive correlations between the proposed model's output and real L2 learner data, particularly for later repetitions and when using specific metrics such as wav2vec 2.0 distortion. Comparative analysis of acoustic perception models suggests that those trained on non-native speech data more effectively simulate L2 learners' perceptual patterns in the proposed simulation model. Based on recent successes of speech representations in various speech tasks such as automatic speech recognition, this work explores state-of-the-art speech representations including wav2vec 2.0, HuBERT, WavLM, and XLS-R for L2 proficiency assessment and L2 pronunciation assessment. Investigation of pronunciation assessment metrics shows that combining repetition with feedback leads to significantly better improvement compared to repetition alone. The integrated computational model and the automatic pronunciation assessment metrics for L2 pronunciation learning contributes to understanding the learning process and has implications for developing more effective teaching strategies and computer-assisted pronunciation training systems.