Robust Speaker Recognition Based On Compressive Sensing

In recent years, speaker recognition as a kind of speech recognition, particularly in the field of biometrics, has drawn increasing attention in the world. It has superiority in practical applications because of its unique convenience, economy and accuracy. As to speaker recognition, feature parameter extraction methods like MFCC are almost perfect when the speech data are pure .However, they will degenerate dramatically in noisy environment. The paper mainly study speaker recognition method in noise environment, and the main work is shown as follow:Firstly, convert noisy speech into Mel spectrum by Mel filter banks, and then judge the condensability of Mel spectrum by discrete haar transform (DHT) of Mel spectrum;Secondly, reconstruct unreliable data using reliable data of Mel spectrum with noise ,then regenerate Mel spectrum by solving the convex optimization nonlinear recovery algorithm;Thirdly, speaker features with auditory model are extracted from reconstructed Mel spectral date, then compare with the speech database template, and obtain recognition results according to a certain similarity criterion.We apply the missing data imputation method based on compressive sensing theory in the front of speaker recognition system, which showed better robustness and higher system recognition rate in a noisy environment.