Text Independent Speaker Verification Based on Mixing ICA Overcomplete Basis Functions

Overcomplete representation of the ICA basis functions is an efficient way to extract the statistical features of signals. By pre-assigned the probability density of the basis coefficients to ensure some statistical characteristic, such as sparseness, we can flexible extract the inner structure of signals. From the point of signal, different ICA basis functions can efficiently represent the local feature information of the signal in different time duration. From the point of the physical feature of speaker, the formants and the changing of formants information represents the physical features and articulation habits of the speaker. In this paper, the ICA overcomplete representation is used to capture the basis functions of the speech signal and the acoustical basis functions of formants, and to build the mixing overcomplete basis functions as the feature information of the speaker. Simulation results based on the proposed approach shows well.