COMPARATIVE EVALUATION OF CLASSIC AND DEEP LEARNING MODELS FOR EARLY DETECTION OF DYSARTHERIA

Abstract

This article analyzes the scientific foundations and practical effectiveness of applying machine learning and deep learning methods for the automatic detection and severity assessment of dysarthria. Dysarthria is a common symptom of neurological disorders such as Parkinson’s disease, cerebrovascular accidents, cerebral palsy, and multiple sclerosis, leading to impairments in articulatory, prosodic, and resonant components of speech. Due to the subjectivity and time consumption of traditional clinical diagnostics, the demand for automated speech processing systems is increasing. In this study, the UA-Speech and TORGO datasets were employed, and classical machine learning models such as RandomForest, XGBoost, LightGBM, SVM, LogisticRegression, and MLPClassifier were compared with deep learning approaches based on CNN architectures. The models were optimized using GridSearchCV, while the pronounced data imbalance was addressed with ADASYN and class weight techniques. Key acoustic features such as MFCC, Mel spectrograms, and pitch were used to construct the feature space. The results indicate that machine learning models can achieve high accuracy in detecting dysarthria symptoms, while hybrid approaches combining CNNs provide improved overall performance. This research lays the groundwork for developing automated early dysarthria diagnostic systems and contributes to the advancement of technologies aimed at clinical practice.