Speech enhancement based on minimum mean-square error short-time spectral estimation and its realization

The paper focuses on theory and realization of speech enhancement based on minimum mean square error short time spectral estimation. It achieves good results well when enhancing speech degraded by stationary additive white Gaussian noise on the condition that only the noisy speech is available. After presenting the ordinary formula of MMSE STSA estimator, its fatal defects in realization especially in real time, are stressed. Meanwhile, another formula of MMSE STSA estimator is promoted, together with a kind of table looking up method for calculating the most time consuming parts in the expression of the estimated value. The new method results in significant amelioration of computing speed, and so can be realized in real time. Finally, a comparison containing several aspects between the three methods discussed in the paper is also included