Selection of speaker independent feature for a speaker verification system

In this paper we propose an optimisation technique to choose a user independent feature subset from the input feature set for a DTW-based text-dependent speaker verification system. The optimisation technique is based on the l-r algorithm, which in essence is the combination of sequential forward and backward search algorithms. The performance criterion used for optimum feature selection is the experimental error rate. The proposed scheme is applied to cepstrum coefficients and their first order orthogonal polynomial coefficients. Experiments are conducted on a data base of 40 Spanish male and female speakers. The results indicate that with the optimised feature set the verification error rate of the system can be improved. Moreover, the speed of verification is significantly increased.