A New Lightweight Dual-Microphone RNN-VAD Assist Adaptive Blocking Matrix for Dual-Channel Speech Enhancement

Abstract

The generalized sidelobe cancellation (GSC) algorithm is widely used in microphone array speech enhancement. However, there are some problems related to it, such as the direction of arrive (DOA) estimation errors, insufficient spatial information utilization, and delay estimation errors, which make the blocking matrix (BM) unable to completely suppress the expected speech, resulting in leakage of the expected speech to the adaptive noise cancellation (ANC) module and the distortion of the desired voice. In this paper, we design a lightweight dual-channel RNN network structure to train the voice activity detection (VAD) which is a key module in GSC. The training process not only uses the amplitude characteristics of the two channels, but also adds the phase difference characteristics. The addition of spatial information can effectively improve the discrimination accuracy of VAD. Then we combine that VAD with the adaptive blocking matrix (ABM) and adaptive noise cancellation (ANC) for noise reduction processing. The simulation results show that the proposed method can significantly reduce the leakage of the desired voice while improving the noise reduction capability of GSC.