Robust Features for Automatic Text-Independent Speaker Recognition Using Ergodic Hidden Markov Models (HMMs)

In this paper, robust feature for text-independent speaker recognition has been explored. Through different experimental studies, it is demonstrated that, these robust features captures speaker specific information effectively by using Ergodic hidden Markov models (HMMs). The study on the effect of feature vector size for good speaker recognition demonstrates that, feature vector size in the range of 18-22 can capture speaker specific related information effectively for a speech signal sampled at 16 kHz, it is established that the proposed speaker recognition system using robust features requires significantly less amount of training data during both in the training as well as in testing. Finally, the speaker recognition studies using robust features for different mixtures components, training and test durations have been exploited. We demonstrate the speaker recognition studies on TIMIT database.