DNN-Based Speech Enhancement via Integrating NMF and CASA

Abstract

In this paper, we come up with a novel speech enhancement method, which integrates nonnegative matrix factorization (NMF) and computational auditory scene analysis (CASA) using deep neural network (DNN). Firstly, we can obtain the basis matrices of speech and noise respectively via NMF and get the ideal ratio mask (IRM) that is based on CASA by using deep neural network. Then, a linear minimum mean square error (LMMSE) filter in fast Fourier transform (FFT) domain is constructed and transformed to the Gammatone domain. Finally, an integrated Wiener-like filter is obtained by combining the filter of NMF with the mask of CASA. By comparing with NMF and CASA methods, the experiments present the superiority of the proposed method.