Prediction of NMF-based Wiener Filter for Speech Enhancement Using Deep Neural Networks
Zhigang Bai, Changchun Bao, Zihao Cui · 2020
In this paper, a novel approach is presented to predict a training target called NMF-based Wiener filter using deep neural networks (DNN) in the nonnegative matrix factorization (NMF) based speech enhancement. The NMF-based Wiener filter, as a masking-based target, is easier than the encoding vectors used in previous algorithms for parameter estimation. The intermediate error of the NMF-based speech enhancement process was reduced due to direct prediction of the NMF-based Wiener filter. The encoding vectors of noisy speech were extracted with the NMF algorithm and normalized to obtain more discriminative input features. The DNN was trained to learn a nonlinear mapping from the encoding vector of noisy speech to the NMF-based Wiener filter. At test stage, the predicted NMF-based Wiener filter was used to enhance noisy speech. The objective evaluations demonstrated that the proposed algorithm outperforms some existing NMF-based and DNN-based methods at various input signal-to-noise ratio (SNR) levels.