A Maximum Likelihood Approach to Masking-based Speech Enhancement Using Deep Neural Network

The minimum mean squared error (MMSE) is usually adopted as the training criterion for speech enhancement based on deep neural network (DNN). In this study, we propose a probabilistic learning framework to optimize the DNN parameter for masking-based speech enhancement. Ideal ratio mask (IRM) is used as the learning target and its prediction error vector at the DNN output is modeled to follow statistically independent generalized Gaussian distribution (GGD). Accordingly, we present a maximum likelihood (ML) approach to DNN parameter optimization. We analyze and discuss the effect of shape parameter of GGD on noise reduction and speech preservation. Experimental results on the TIMIT corpus show the proposed ML-based learning approach can achieve consistent improvements over MMSE-based DNN learning on all evaluation metrics. Less speech distortion is observed in ML-based approach especially for high frequency units than MMSE-based approach.