Generalized Spatio-Temporal RNN Beamformer for Target Speech Separation

• DOI: 10.21437/interspeech.2021-430
• Publication Date: 2021-08-27T05:59:39Z
• Authors: Yong Xu; Zhuohuang Zhang; Meng Yu; Shi-Xiong Zhang; Dong Yu
• Keywords: Separation (statistics); Source Separation

Although the conventional mask-based minimum variance distortionless response (MVDR) could reduce non-linear distortion, residual noise level of MVDR separated speech is still high.In this paper, we propose a spatio-temporal recurrent neural network based beamformer (RNN-BF) for target separation.This new beamforming framework directly learns weights from estimated and spatial covariance matrices.Leveraging on temporal modeling capability RNNs, RNN-BF automatically accumulate statistics matrices to learn frame-level in recursive way.An RNN-based generalized eigenvalue (RNN-GEV) more RNN (GRNN-BF) are proposed.We further improve RNN-GEV GRNN-BF by using layer normalization replace commonly used mask matrices.The proposed obtains better performance against prior arts terms quality (PESQ), speech-to-noise ratio (SNR) word error rate (WER).