Speech Enhancement from Fused Features Based on Deep Neural Network and Gated Recurrent Unit Network

Abstract

\n Speech is easily interfered by the external environment in reality, which will lose the important features. Deep learning method has become the mainstream method of speech enhancement because of its superior potential in complex nonlinear mapping problems. However, there are some problems are exist such as the deficiency for the learning the important information from previous time steps and long-term event dependencies. Due to the lack of the correlation in the same layer of Deep Neural Networks (DNNs), which is an existing typical intelligent deep model of speech signal, it is difficult to capture the long-term dependence between the time-series data. To overcome this problem, we propose a novel speech enhancement method from fused features based on deep neural network and gated recurrent unit network. The method takes advantage of both deep neural network and recurrent neural network to reduce the number of parameters and simultaneously improve speech quality and intelligibility. Firstly, DNN with multiple hidden layers is used to learn the mapping relationship between the logarithmic power spectrum (LPS) features of noisy speech and clean speech. Secondly, the LPS feature of the deep neural network is fused with the noisy speech as the input of gated recurrent unit (GRU) network to compensate the missing context information. Finally, GRU network is performed to learn the mapping relationship between LPS features and log power spectrum features of clean speech spectrum. Experimental results demonstrate that the PESQ, SSNR and STOI of the proposed algorithm are improved by 30.72%, 39.84% and 5.53% respectively compared with the noise signal under the condition of matched noise. Under the condition of unmatched noise, the PESQ and STOI of the algorithm are improved by 23.8% and 37.36% respectively. The advantage of the proposed method is that it uses of the key information of features to suppress noise in both matched and unmatched noise cases and the proposed method outperforms other common methods in speech enhancement.