Online Speech Dereverberation Using RLS-WPE Based on a Full Spatial Correlation Matrix Integrated in a Speech Enhancement System

Reverberation may degrade the performance of automatic speech recognition significantly. Although the weighted prediction error (WPE) provided impressive results, most of the conventional WPE methods were based on batch processing. An efficient WPE method based on frame-by-frame adaptation of the prediction filter in the recursive least squares (RLS) framework was proposed, but the spatial correlation matrix was oversimplified to be a scaled identity matrix. In this paper, we derive a generalized online RLS-WPE method allowing a full spatial correlation matrix. Experimental results showed that the method using a diagonal matrix achieved better performance than that using a scaled identity matrix and employing a full matrix further improved the performance. Furthermore, we also integrated the algorithm into a speech enhancement system, including stereo adaptive echo cancellation and minimum variance distortion response beamforming, implemented onto Xilinx Zynq Ultrascale+ MPSoC.