Deep Neural Network Based Multi-Channel Speech Enhancement for Real-Time Voice Communication Using Smartphones

Recently, the performance of speech enhancement has been improved via deep neural networks. However, most of them are too heavy for voice communication using smartphones, and some are non-causal systems. In this paper, we introduce some effective techniques improving the performance even with light-weight models at causal system. We extract the input features by incorporating two kinds of beamformers. Furthermore, a normalization scheme is proposed to diminish the inter-channel variance between two beamformer outputs. The experimental results show the superiority of the proposed features. Moreover, the proposed method is extendable to any number of microphone systems without additional model training.