Optimized Variable Size Windowing Based Speaker Verification

Abstract

In recent years the variances of speech features of speaker verification system were measured by computing covariance matrix parameterized through its eigenvalues and vectors by keeping fixed sliding window size. The computed eigenvectors were weighted with its corresponding magnitude and normalized. Then, the features were extracted and fused using different fusion techniques for recognizing the speaker. However, this approach was not suitable for all types of datasets and some significant feature information may be lost during extraction based on fixed window size. Hence in this article, the variable size sliding window is applied for Speaker Verification system. Initially, the speech signal is considered as input and the FMPM features are extracted using FDLP, MHEC and PNCC including MFCC based on the variable size of a sliding window. Here, the sliding window size is optimized by Modified Grey Wolf Optimization (MGWO) algorithm which is also used for selecting the classifier parameters and most optimal features adaptively. The most optimal features are selected from the extracted FMPM and classified by using GMM classification. Thus, the proposed approach allows continuous adaptation of SV using variable window size and classifier parameters. Finally, the considerable improvements in Speaker Verification are observed through experimental results.