A novel single channel speech enhancement based on joint Deep Neural Network and Wiener Filter

In this paper, we present a novel single channel speech enhancement method based on joint Deep Neural Network (DNN) and Wiener Filter as a whole network named Wiener Deep Neural Network (WDNN). The proposed method contains two stages: the training stage and the enhancement stage. In the training stage, WDNN predicts the clean speech magnitude spectra and the noise magnitude spectra from noisy speech features simultaneously. Then, the Wiener filter is placed on top of the two output of the neural network as an extra layer to generate the enhanced speech magnitude spectra. Finally, we use the phase of noisy speech to reconstruct clean speech. In the enhancement stage, the well-trained WDNN is fed with the features of noisy speech in order to obtain the enhanced speech. Extensive experimental results show that the proposed method outperforms state-of-the-art methods such as the non-negative matrix factorization (NMF) and the tradition DNN methods.