Deep neural network based monaural speech enhancement with sparse and low-rank decomposition
Wenhua Shi, Xiongwei Zhang, Meng Sun, Xia Zou, Yanmin Wei, Gang Min · 2017
Taking into account of the sparse and low-rank structure of noisy speech spectrogram and good temporal-spectral preservation characteristics of deep neural network (DNN), a novel monaural speech enhancement framework combining DNN with sparse and low-rank matrix analysis is proposed in this paper. Sparse representation of target speech is obtained via sparse and low-rank decomposition on the noisy speech spectrogram. DNN is used to estimate the non-linear function which maps the sparse and the magnitude spectrum features of the noisy speech to the magnitude spectrum of the target speech. Network parameters are optimized by back-propagating the reconstruction error using MMSE criterion. Evaluations in terms of PESQ and LSD show that the proposed method is superior to the supervised NMF and the regression DNN based speech enhancement method.