Improved Sparse NMF based Speech Enhancement Method with Deep Neural Network

Abstract

Abstract. Considering the sparsity characteristic of speech signal in time-frequency domain and the non-linear model ability of deep neural network, an improved sparse non-negative matrix factorization based speech enhancement method is presented in this paper. Deep neural network is employed to learn the sparse encoding coefficients of speech and noise from noisy observation. The estimated clean speech is obtained by applying the wiener filter on the magnitude spectrogram of noisy speech. The experimental results show the superiority of proposed method under stationary and non-stationary conditions.