Key points are not available for this paper at this time.
Abstract In this paper, a multi-channel speech enhancement structure combining beamformers and lightweightneural networks is proposed. By using a first-order differential beamformer to improve the signalto-noise ratio of the target speech, the burden on the subsequent neural network is reduced, thuslowering the complexity of the required neural network. A two-stage gated recurrent neural networkis employed, where the first stage recurrent neural network processes the channel characteristics ofthe speech and the second handles the frequency domain features of the speech. The frequency banddivision combined with convolution is used to further compress the network scale. With suitable lossfunction, the speech enhancement performance of the model is further improved.The proposed modelstructure is trained, evaluated, and validated using simulated multi-channel microphone array datasetsgenerated from the public TIMIT dataset. The results demonstrate that our model achieves goodmulti-channel speech enhancement performance with relatively small parameter and computationalrequirements when compared to popular existing approaches.
Cao et al. (Tue,) studied this question.