Signal representation serves as an integral part of signal processing algorithms. The selection of appropriate representation for speech signals plays an important role in speaker recognition or identification applications. In this paper, an efficient representation of signal features based on Dual Tree M-band Wavelet Transform (DTMWT) is presented and these features are utilized for speaker recognition. DTMWT is used for the analysis of speech signals as they are absolute shift invariant than wavelets. The performances of signal features based on DTMWT are analyzed with the help of Gaussian Mixture Model (GMM) classification system with various Gaussian components. Experimental results on CHAINS corpus show that the DTMWT-GMM system provides an accuracy of 98.61% for 8-speaker set, 97.22% for 16-speaker set, and 96.87% for 32-speaker set on a closed-set identification.
Volume 10 | 11-Special Issue
Pages: 369-377