Multichannel Speech Enhancement in Vehicle Environment Based on Interchannel Attention Mechanism

Figure 3

The structure of the TsBiLSTM module and illustration of the proposed module architecture. (a) The input of the TsBiLSTM includes interchannel features and original waveform. (b) The processing chain shows the two-stage BiLSTM with residual connections. (c) Carry out 1 × 1 convolution on the output after overlap-add method operation, then operate conv1d layer with sigmoid and tanh activation function, respectively, and obtain beamforming filter.