Abstract: An adaptive speech enhancement method based on Data-Driven Dictionary and overcompletely sparse representation theory is proposed. Firstly, using the K-singular value decomposition (K-SVD) algorithm, a dictionary that describes the clean speech content effectively is trained. Secondly, the prime threshold is adaptively selected according to noise variance of original noisy speech signal and the speech signal’s sparsest coefficient vector is obtained through Orthogonal Matching Pursuit algorithm. And then the speech signal is recovered and speech enhancement is achieved. Different from the conventional techniques which improve the speech signal quality by suppression of noise and reduction of distortion, we select the appropriate atoms to represent speech signal. Thus, clean signal is separated from the noisy speech signal. In white or colored noise interference, the reconstructed speech signal via the proposed algorithm is evaluated by the objective and subjective evaluation. The experimental results show that the proposed algorithm can get ride of the addictive noise and improve speech quality.