Recently, proposed recurrent neural network (RNN) based all deep learningminimum variance distortionless response (ADL-MVDR) beamformer method yielded superior performance over the conventional MVDR by replacing the matrixinversion and eigenvalue decomposition with two RNNs . Temporal-spatial self-attention module is proposed to better learn thebeamforming weights from the speech and noise spatial spatial spatial covariance matrices . Thetemporal self-association module could help RNN to learn global statistics ofcovariance matrics . The spatial self-appearance module is designed to attend on the cross-channel correlation in the covariance matrix matrices. Furthermore, amulti-channel input with multi-speaker directional features and multi

Author(s) : Xiyun Li, Yong Xu, Meng Yu, Shi-Xiong Zhang, Jiaming Xu, Bo Xu, Dong Yu

Links : PDF - Abstract

Code :
Coursera

Keywords : spatial - multi - module - rnn - speech -

Leave a Reply

Your email address will not be published. Required fields are marked *