A Neural Beamforming Network for B-Format 3D Speech Enhancement and Recognition

Multi-channel speech enhancement is gaining increasing interest in recent years. By combining the beamforming framework with the deep neural network, significant improvement on speech enhancement performance has been achieved. While the neural beamformers designed for the distributed microphone arrays are deployed for practical applications such as teleconferencing and surveillance, there is less approach designed for the co-located microphone arrays such as the soundfield microphones. In this work, a new neural beamforming network is proposed for B-format 3D multi-channel speech enhancement and recognition. The proposed method incorporates the traditional beamforming structure with the deep neural network specifically for the B-format channels (the first-order Ambisonics). The proposed method has ranked the 1st place of the 3D Speech Enhancement task in the MLSP L3DAS21 Challenge while significantly outperformed the baseline system on the WER and STOI metrics.