Best Speaker-Based Structure Tree for Speaker Verification

C. Tadj, C. Gargour and N. Badri

In this paper we study the use of the Wavelet transform for text-dependent speaker verification purposes. A new algorithm to construct the best admissible tree is proposed which has been used to obtain a speaker dependent tree library. Every tree in this library corresponds to the best structure for a given speaker, therefore the extracted parameters from a given tree are well suited and discriminative to the considered speaker and hence appropriate for speaker verification purposes. Experiments have been conducted using data extracted from the Yoho database. The results that have been obtained show a good level of efficiency and robustness compared to those which can be obtained by using MFCC or other wavelets and wavelets-based methods.