Deep neural network based speech enhancement using mono channel mask

Abstract

Getting enhanced speech from the noisy speech signal is a task of particular importance in the area of speech processing. Here we propose a deep neural network (DNN) based speech enhancement method utilising mono channel mask. The proposed method employs cochleagram to find an initial binary mask. Then modified sub-harmonic summation algorithm is applied on initial binary mask to obtain an intermediate mask. The spectro-temporal features of this intermediate mask are fed to DNN. DNN finds out the correct spectral structure in the frames associated with the target speech which are further used to develop the mono channel mask. Speech signal is reconstructed using mono channel mask. Mono channel mask avoids the unnecessary interference from the noisy time–frequency (T–F) units. Objective evaluations done using perceptual evaluation of speech quality (PESQ) and normalized source to distortion ratio indicate that the proposed method outperforms the state of the art methods in the area of speech enhancement. Obtained values of PESQ shows that proposed method improves the quality of the speech in noisy conditions. The experimental results present the effectiveness of the mono channel mask in speech enhancement. The proposed method gives better performance compared to other methods.