Total variability factor analysis for dysphonia detection

Abstract

This paper proposes the use of a new feature extraction technique named I-Vectors to detect and classify Dysphonic voice data. After initial pre-processing and framing, a total of 52 important acoustic features are extracted from the speech. Gaussian Mixture Model (GMM) algorithm makes use of an Expectation–Maximization (EM) algorithm to compute the zeroth, first and second order statistics pertaining to the UBM model for final extraction of I-Vectors. This high dimensional supervector is transformed into a low dimensional subspace called the Total Variability space (TV subspace) following which a new vector set (called I-Vectors) in extracted using the UBM model and Baum-Welch statistics. A total of 110 I-Vectors are extracted for the given input data after undergoing Cholesky whitening. This newly created data is now used for classifying normal and dysphonia samples. The application of Support Vector Machines (SVM) along with the extracted I-Vectors gives 98% accuracy on the unknown data set, which is a significant improvement in accuracy, as compared to 92% obtained without using I-Vectors.