Comparative Analysis of Machine Learning Models for Speech Disorder Detection

Speech disorders like dysarthric syndrome, stuttering, and speech sound disorder in children are defined as significant impediments to integration into the social world as well as to education and quality of life. The traditional evaluation methods are mainly based on auditory perceptual analysis and are known to have inter-rater and resource variability issues. Recent developments in machine learning (ML) and computational intelligence (CI) break these earlier limitations and open their way to automated identification of speech disorder detection and assistive applications. This paper reports on key methodologies applied, in particular landmark analysis, CNNs, and anomaly detection methods, for the screening of disordered speech in children and the support of adults with acquired neuromotor impairment. The capacity of strategies to promote earlier diagnosis, enhance the performance of Automatic Speech Recognition (ASR) for individuals suffering from speech disorders, and facilitate easier intervention. Our comparative analysis of five ML models—LDA, CNN, Wave2Vec, SVM, and BiLSTM—reveals that Wave2Vec outperforms the others in terms of accuracy and robustness in speech disorder detection.