Abstract: FIR adaptive filtering algorithm can improve the adaptive speed based on spectral subtraction and LMS of the short-term non-stationary or wide range signals for speech enhancement in condition broadband noise. This algorithm detects an expected speech signal of derived from short-term noise with spectral subtraction, this expected signal is used as reference signal of the output filter. The method improves performance of filter and chooses differences as the error signal between the referenced signal and the output filter. The filters of reconstruction are revised directly by filter corrective value based on LMS of weigh. This algorithm adopts the way of normalization LMS and symbolization LMS and blocking LMS in the steepest descent adjusting the weights, and simplifies step selection of converge weight, and reduces the computation of corrective weight. The simulation experiment shows that the algorithm can effectively enhance speech of all kinds of SNR, and it is superior to spectral subtraction. The result shows that the enhancement method is satisfaction in the 3 dB.