Abstract: In order to solve the problems such as low recognition rate and poor anti-interference ability of speech signal feature extraction, a method of feature extraction based on improved empirical modal decomposition algorithm is presented. Classification by the method including noise speech signal decomposition, two types of modal component processing, reconstruction and feature extraction, respectively, to solve present most speech signal feature extraction process will filter out part of the original signal, on the basis of effectively eliminate the noise signal, as much as possible to save the original signal. And the recognition performance of system is improved obviously. Experimental results show that the proposed algorithm can achieve a 95. 5% recognition rate without adding noise. Compared with several traditional algorithms, this algorithm maintains a high recognition rate when adding different proportion of noise.

Key words: noisy speech signal, mode decomposition, feature extraction, reconsitution