Allocate observation numbers adaptively

according to sparsity in wavelet tree of speech frames

change number of wavelet tree nodes with different observation numbers. This method solves mismatching problem between the nodes number in the tree model and measurement of speech signal. Denoising the preprocessed speech signal by improved wavelet

then

through Gammatone filters to deal with the enhanced speech signal

extract feature parameters GFCC. Simulation experiment results demonstrate that SNR increases 14 % compared with traditional wavelet method

effectively reduce effect of noise in speech signal and the system recognition rate and robustness are improved obviously.