Vocal Test Analysis for the Assessment of Adductor-type Spasmodic Dysphonia

Adductor-type Spasmodic Dysphonia is a task-specific focal dystonia characterized by vocal folds' adductor spasms. These involuntary contractions interrupt speech causing strain and strangled voice breaks. The purpose of this paper to is to develop a robust machine learning approach to detect spasmodic dysphonia from voice samples, using balanced data, 10-fold cross validation, and thorough feature selection method based on the Genetic Algorithm. The voice features were analysed using different classifiers such as Naïve-Bayes, Multi-Layer Perceptron, Support Vector Machine, and Random Forest. Statistical analysis was applied to test for significance and superior performance. Results showed that sustained phonation provide higher accuracy models. In addition, Naïve-Bayes outperformed all classifiers with a maximum of 100% and an average of 98.33%. The Genetic Algorithm wrapper feature selection method proved to provide superior performing features than previous researches.