Robust Speaker Recognition Based On Compressive Sensing

Speaker recognition has been highly anticipated in the field of biometric authentication becauseof its unique advantages. MFCC parameter which is most commonly used has a very goodrecognition results in the absence of noise, but there are still two major problems:(1) The amount ofparameters is very large;(2) In a noisy environment, the system recognition rate significantlydecreased. This thesis uses the emerging compressed sensing technology to improve the traditionalMFCC parameters, and presents CS-MFCC parameter. CS-MFCC parameters make storagecapacity decreased to1/n of the original (n is the compression ratio of the observing matrix), Onthis basis, mainly to complete the following tasks:(1) Proposed a new parameter based on line ladder matrix and prove when compression ratio ofmatrix is4can identify the best. In noise-free environment, the Fish ratio of the newparameter based on line ladder matrix is higher than MFCC parameters, the recognition rateincreased significantly. While the new parameter based on line ladder matrix can weakenimpulse noise and Gaussian noise, the robustness of the system increased.(2) Proposed a new recognition parameters based on weighted cycle matix, it performsnon-uniform sampling and enhances low frequency sampling through adjusting themeasurement matrix coefficients, thus suppressed high-frequency noise. By matlabsimulation proves the robustness of the system is significantly improved.