Eigenanalysis applied to speaker identification using gammatone auditory filterbank and independent component analysis

In this paper, we present a novel algorithm specifically designed for speaker identification applications. The proposed algorithm is an extension of the previously developed Gammatone auditory filterbank and independent component analysis (GTF-ICA) method. GTF-ICA extracts features which emphasis the difference in the statistical structures in human cochlea frequency bands among the speakers. This method results in significant improvements in identification accuracy over conventional methods when speech is corrupted by additive noise or the training and testing environments are mismatched. However GTF-ICA introduces high computational cost during identifying the speaker. In our proposed algorithm, eigenanalysis is applied to the extracted features to capture the most important features. Therefore the proposed algorithm reduces the dimension of the feature so as to decrease the computation load. In addition, it increases the identification rate in noisy environment since the noise elements are deleted by the eigenanalysis process.