Improved Text-Independent Speaker Identification and Verification with Gaussian Mixture Models

Abstract

Speaker recognition systems proposed in the literature are conditioned by many factors such as the features used, the type of data to be processed, data duration and the approach to be used. This work presents a novel speaker recognition application using a new dimensional cepstral feature vector for building Gaussian Mixture Models (GMM) for speaker identification and verification systems. Recognition results obtained with the proposed system are compared to the state of the art systems requiring high dimensional feature vectors. We experimentally evaluate this work with speakers taken from TIMIT database. The new system gave substantial improvements even when trained with limited data, which explain the effectiveness of the new feature vectors with this approach which can be comparable to the use additional algorithms or approaches.