Robust speaker identification in noisy environment using cross diagonal GTF-ICA feature

In this paper, we present a novel feature specifically designed for speaker identification in noisy environments. Gammatone auditory filterbank and independent component analysis (GTF-ICA) feature emphasises the difference in the statistical structures of the human cochlea frequency bands among speakers. This method results in significant improvements in the identification accuracy over the conventional methods when speech is corrupted by additive noise and when the training and testing environments are mismatched. However GTF-ICA imposes high computation cost in comparison to the other features. The proposed feature alleviates the computation by considering the cross diagonal elements of GTF-ICA feature matrix used with the Gaussian mixture models (GMM). The proposed algorithm is more computationally efficient than our pervious one. In addition, it increases the identification rate in noisy and mismatched environments.