AR-vector using CMS for robust text independent speaker verification

Abstract

This paper presents the performance of the AR-vector with cepstral mean subtraction (CMS) used to compensate the distortions caused by distinct telephone channels. The speaker recognition performance obtained with the use of CMS is compared with a system without compensation. With 60 s of speech signal used for training and 30 s used for testing, the error rate without channel normalization is around 2.82% against the 1.65% achieved with CMS. For 10 s testing time, the error rate dropped from 5.40% to 3.80% when using CMS. For the lowest testing time (3 s), the error rate of the AR-vector is close to 19% regardless of whether or not the normalization technique is used. Although there is a clear improvement in performance when using CMS, it is not of major significance. This leads to the conclusion that the AR-vector classification system is somewhat robust to channel distortion, especially as the testing time decreases.