Dimensionality reduction for text-independent speaker identification using Gaussian mixture model

Reducing the dimensionality of the training and testing data is crucial for text-independent speaker identification tasks. In this paper, the performance of various dimensionality reduction techniques is evaluated for speaker identification systems using Gaussian mixture model (GMM) as the statistical classifier. An enhancement of the standard linear discriminant analysis (LDA) is proposed in which class distributions are assumed to follow Gaussian mixture distribution. This assumption is more appropriate for asymmetric and multimodal class conditional densities. In addition, a new feature selection technique based on the QR factorization method is introduced. Computer simulation results reveal that the proposed modification to the LDA outperforms the standard algorithm in terms of classification accuracy. Moreover, the QR-based selection technique produces comparable results to other prominent dimensionality reduction techniques