Incorporating Group Update For Speech Enhancement Based On Convolutional Gated Recurrent Network

To further improve the performance of speech enhancement methods based on deep neural networks, this paper proposes a speech enhancement network that can make full use of noisy speech characteristics in the time-frequency domain. First, based on the local correlation of noisy speech in the time-frequency domain and the spatial structure of frequency features, by using a recurrent neural network (RNN) to model the time correlation of noisy speech, and using a convolutional neural network (CNN) to calculate the frequency features of noisy speech, a convolutional gated recurrent network (CGRN) is built for speech enhancement. Second, based on the different variation characteristics of noisy speech over time, a group update mechanism is introduced to further improve CGRN; by artificially dividing the hidden layer features of the recurrent neural network in CGRN into three groups and updating them in three different ways, the CGRN incorporating group update (CGRN-GU) divides the variation characteristics roughly into three cases and can better track the changes of noisy speech over time. Finally, a causal speech enhancement method is proposed using the convolutional gated recurrent network incorporating group update, and extensive experiments are conducted on a public dataset. The experimental results show that, in the comparison with the state-of-the-art methods, the speech enhancement method proposed in this paper has better speech enhancement performance than other causal methods, and the proposed causal speech enhancement method even outperforms most non-causal methods.