Abstract: This paper proposes an improved Nonnegative Matrix Factorization based speech enhancement algorithm. It mainly consists of a training stage and an enhancement stage. Firstly, to reduce the complexity in the training stage, only noise dictionary is trained using Convolution Nonnegative Matrix Factorization as the prior information. During the enhancement stage, considering the different sparse properties shown between noise and speech, the speech is separated from noise by using Sparse Nonnegative Matrix Factorization. Particularly, the update process is solved by putting the theory of Alternating Direction Method of Multipliers, which can solve the problems such as poor local optima and limitation of convergence to zero in the traditional speech enhancement using Nonnegative Matrix Factorization. Finally, to take advantage of the unsupervised speech enhancement algorithm, Spectral Subtraction and Minimum Mean Square Log-spectral based methods are combined to improve the performance. Extensive experiments indicate that the proposed algorithm shows promising noise suppression ability under ten various noise conditions.