Artificial Neural Network and SVM based Voice Disorder Classification

In cognitive infocommunication inspired speech processing one of the most important, but challenging areas of research is the analysis and classification of voice disorders. This paper investigates and compares various input vectors and classification models to find the use of input vectors for separating healthy from disordered voices. We compare acoustic parameters as input vectors with phone-level posterior probabilities computed by the DNN soft-max layer of the speech recognition system. An attempt is made to separate voice disorders from healthy voice in adults using different classification methods. The classification is implemented using Support Vector Machine (SVM) and Sequential Deep Neural Network (DNN). Our results show that using acoustic parameters instead of phone-specific posteriors as input features increases the accuracy of the classification, furthermore the DNN approach outperformed the SVM classifier in case of leave-one-out cross validation (LOOCV) and in case of 70-30% data split method as well.