Abstract: An improved speech enhancement method based on sparse representation of spectral features is proposed from the perspective of dual dictionary learning,noise power spectrum estimation and speech amplitude spectrum reconstruction.In the dictionary learning stage,the power spectrum and amplitude spectrum features are combined,and the discriminative dictionary learning method is used to reduce the coherence of the speech dictionary and the noise dictionary.In the speech enhancement stage,a noise power spectrum estimation method is proposed to track and estimate the non-stationary noise.Considering the inconsistency between the amplitude spectrum and the power spectrum characteristics for different noises,a speech reconstruction weight table is designed.The two way signals recovered from the amplitude spectrum and the power spectrum are adaptively weighted,and the enhanced speech signal is obtained by combining the phase compensation function.As shown in experimental results that,compared with the single spectral feature speech enhancement method,this method improves the average of 31.6% in stationary and non-stationary noise environments,improving the performance of the speech enhancement method.