A Machine Learning Dyslalia Mispronunciation Detection Framework for Interactive Speech Therapy

Speech disorders, such as dyslalia, are highly prevalent among children. Traditional speech therapy methods are constrained by limited accessibility, and low engagement. In this paper, we investigate the use of machine learning (ML) in enabling real-time pronunciation practice and feedback. A synthesized dataset that includes both correct and dyslalia-affected pronunciations was developed. Our ML framework incorporates comprehensive preprocessing and feature extraction using Mel-Frequency Cepstral Coefficients (MFCC). Multiple classifiers were evaluated, and three training approaches were explored. To demonstrate the practical utility of the system, we integrated the framework into a virtual reality (VR) game, LingoLand, which provides engaging speech therapy for children. The classification results achieved accuracies ranging from 83% to 95 %, with the best performance observed using the support vector machine (SVM) and XGBoost.