This paper mainly discusses the speech enhancement algorithm based on the Minimum Mean Square Error（MMSE） estimation of speech short time spectrum. On the basis of the general MMSE estimation

a new algorithm using interframe frequency domain distribution constraint in MMSE estimation is proposed

if takes into account the correlations between interframe spectrum and incorporates them into the estimation.Due to the fact that the distribution of speech spectrum is unknown

two solutions are adopted: Gaussian distribution assumption and statistic approximation obtained by real speechmaterials.The new algorithm has been evaluated in our speech enhancement experimental system

and compared with other algorithms. The results show that the new algorithm is more effective to reduce the background noise without significant speech degradation.