Speaker Verification using 3-D ROC Curves for Increasing Imposter Rejections

A speaker verification system (SVS) is developed based on a bank of radial basis function (RBF) neural modules (BRBFNM). The output thresholds of the RBF networks are set using 3-dimensional receiver operating characteristic (ROC) curves. Moreover, a customized cepstral-based feature extraction pre-processing approach is used at each module to select only those features of the speakers that will contribute to enhancing the SVS's performance. These feature vectors are used to train and test the BRBFNM. Each of four speakers was asked to speak 10 different key words 12 times. From the 479 signals (one speaker missed one word), 279 signals were used for training and 200 for testing. The BRBFNM can achieve a 90.5% correct verification rate (CVR), thus reducing false acceptances (i.e., increasing imposter rejections).