Speaker Recognition is important in the field of computer intelligent interface and human-computer interaction, the task of which is to analyze the speech wave of speakers, model them with the extracted speech features and recognize their identities.Recognition performances of traditional feature extraction methods like MFCC are almost perfect when the speech data are clean. However, they degenerate dramatically in noisy environment.In this paper, a new approach called sparse discriminant analysis which imposes sparse constraint on the linear discriminant analysis is developed and its corresponding algorithm derived by gradient descent method is also given. Furthermore, their generalized tensor based method and algorithm are also explored.Experiment results demonstrate that our methods not only achieve perfect recognition performances with clean speech data, but also improve the performances in noisy environment. The good results are based on the fact that our methods combine both the discriminant power of the linear discriminant analysis and the noisy removing power of the sparsity. |