Performance-Based Comparative Analysis of Machine Learning Algorithms for English Pronunciation Assessment Systems
Abstract
English has become a dominant medium of international communication, education, science, and digital interaction, which has increased the number of learners who need objective and scalable pronunciation support. Traditional teacher-based pronunciation evaluation is often time-consuming and subjective; therefore, automatic pronunciation assessment systems are increasingly used to provide faster feedback and improve learning efficiency. This study presents a performance-based comparative analysis of machine learning algorithms used in English pronunciation assessment systems. A pilot numerical dataset was constructed using pronunciation-related acoustic features, including MFCC distance, pitch deviation, duration error, formant error, energy deviation, stress error, rhythm deviation, and Goodness of Pronunciation score. Four algorithmic approaches were evaluated: Support Vector Machine, Random Forest, 1D Convolutional Neural Network, and an HMM-GOP baseline. The models were trained and tested under the same experimental conditions, and their effectiveness was measured using accuracy, precision, recall, F1-score, ROC-AUC, training time, and testing time. The experimental comparison showed that feature-based classical and ensemble models provide stable results on small numerical datasets, while the 1D-CNN requires larger or more structured speech representations to reach its full potential. The HMM-GOP baseline demonstrated strong performance in phoneme-level pronunciation scoring due to its direct use of pronunciation likelihood information. The findings indicate that the choice of algorithm should be aligned with dataset size, feature structure, and assessment objectives. The study provides practical guidance for developing efficient English pronunciation assessment systems for language learners.