Voice Pathology Detection and Multi-classification Using Machine Learning Classifiers

Abstract

Automatic voice pathology detection can provide objective estimation and prevention in the early stages of voice diseases. A voice pathology detection and multi-classification method using an audio feature set extracted from glottal flow waveform is proposed in this work. In addition, a feature selection method of Fisher discrimination criterion is used to screen more valuable features and eliminate the features with low efficiency. The discrimination ability and effectiveness of the selected features were verified using three dimensional scatter plot and box plot in this study. One of the contributions of this paper is to investigate and evaluate the performance of different machine learning classifiers value for voice pathology detection and multi-classification. All experiments were carried out using the Massachusetts eye and ear infirmary database. Accuracy, sensitivity, specificity and receiver operating characteristic area are used as evaluation indexes to compare the performance of different machine learning classifiers. Each machine learning classifier had a good performance using the proposed feature set with nearly 100% accuracy in detection and higher than 90% accuracy in multi-classification. The experimental results indicate that the proposed method is valuable for voice pathology detection and multi-classification, especially in multi-classification.