Abstract

In contrast to the minimum mean square error (MMSE)-based noise cancelation techniques, we propose a method to enhance speech by means of finding a mapping function between noisy signal and clean speech signals based on deep neural networks (DNNs). In order to handle a wide verity range of additive noises in real-world scenarios, a large number of training set that contains many possible combinations of speech and noise types, is first designed. The DNN architecture is then employed as a nonlinear regression function to ensure a powerful modeling capability. To further improve the DNN-based speech enhancement system we proposed a technique called global variance equalization to remove the over-smoothing problem of the regression model to improve the generalization capability of DNNs. To further improve the quality of enhanced speech and generalization capability of DNNs. First, equalization between the global variance (GV) of the enhanced features and the reference clean speech features is proposed to alleviate the over- smoothing issue in DNN- based speech enhancement system.