EMBEDDING AND BEAMFORMING: ALL-NEURAL CAUSAL BEAMFORMER FOR MULTICHANNEL SPEECH ENHANCEMENT

Andong Li, Wenzhe Liu, Chengshi Zheng, Xiaodong Li

Abstract: Standing upon the intersection of traditional beamformers and deep neural networks, we propose a causal neural beamformer paradigm called Embedding and Beamforming, and two core modules are designed accordingly, namely EM and BM. For EM, instead of estimating spatial covariance matrix explicitly, a 3-D embedding tensor is learned with the network, where both spatial-spectral-temporal discriminative information can be represented. For BM, a network is directly leveraged to derive the beamforming weights so as to implement filter-and-sum operation. To further improve the speech quality, a post-processing module is introduced to further suppress the residual noise. Based on the DNS-Challenge dataset, we conduct the experiments for multichannel speech enhancement and the results show that the proposed system outperforms previous advanced baselines by a large margin in multiple evaluation metrics.


EaBNet Architecture:



Samples:

Noisy                                                               CTSNet                                                       GaGNet                                                      

FasNet+TAC                                                    MC-ConvTasNet                                         MIMO-UNet                                                      

MB-MVDR(oracle)                                           EaBNet(Pro.)                                             EaBNet+PostNet(Pro.)                                                  

Utterance 1 (SNR=-5dB, target-doa=15°, inter-doa=30°)

                          

                                       

                          

Utterance 2 (SNR=-2dB, tar-doa=10°, inter-doa=110°)

                          

                                       

                          

Utterance 3 (SNR=0dB, tar-doa=30°, inter-doa=85°)

                          

                                       

                          



Experimental Results: