Speaker verification based on a vector quantization approach that incorporates speaker cohort models and a linear discriminator

This report describes a speaker verification system based on a vector quantization (VQ) approach that incorporates dynamic time warping (DTW), cohort models, and a discriminator to separate the true speakers and imposters. This system is designed for telephone network applications and provides good performance for different telephone handsets and network conditions. The speaker verification system is text dependent, and each subscriber uses a personal pass phrase to verify his/her identity. A distortion vector is computed using VQ encoding and DTW distortions with respect to speaker-dependent and cohort code books. The components of this distortion vector are applied to a linear discriminator to validate the identity of the speaker. A speaker verification database was used to evaluate the performance of this approach, and it was found that the performance is significantly better compared with that of a basic VQ system. The system achieved an equal error rate (EER) of 0.92% when the true speaker and the imposters spoke different pass phrases and 4.30% when they spoke the same phrase.