Dual-Masking Wind Noise Reduction System Based on Recurrent Neural Network

In this paper, we adopt the architecture of permutation invariant training (PIT) model. We take advantage of the dual mask features of the speech separation architecture and combine the results of the two masks to synthesize a better signal with a specific ratio. We use bidirectional gated recurrent unit (BGRU) to find appropriate weights for the features after short time Fourier transform (STFT). A mask finds the signal you want to keep. Another mask finds the unwanted signals. Compared with the traditional method for eliminating wind noise, our proposed method can achieve better noise reduction for non-stationary and non-periodic wind noise.