Abstract: In traditional speech enhancement algorithm the speech spectral amplitude is assumed to be mutually independent. Little work has been done to incorporate the time and frequency dependencies of speech. Without exploring the structure information of the time and frequency neighbors limit the performance of traditional speech enhancement algorithms. In this paper, we propose a novel speech enhancement algorithm based on data field theory, which is capable of modeling the time and frequency dependencies of speech. Data field defines the distribution of the magnitude of speech spectral samples conditioned on the values of their time and frequency neighbors. This formulation allows the explicit inclusion in the amplitude estimation model of both time and frequency dependencies that exist among the amplitudes of speech spectral. The proposed algorithm is evaluated by applying it to enhance noisy speech at various noise levels. Systematic evaluation shows that the proposed algorithm offers improved signal to noise ratio and presents an enhanced ability in preserving the weaker speech spectral components compared to Martin’s algorithm.