Robust SBR method for adverse Mandarin speech recognition

Abstract

An RNN-based robust signal bias removal (RRSBR) method is proposed for improving both the recognition performance and the computational efficiency of the SBR method fbr adverse Mandarin speech recognition. It differs from the SBR method in using three broad-class sub-codebooks to encode the feature vector of each frame and combining the three encoding residuals to form the frame-level signal bias estimate. A novel approach involving softly combining the board-class encoding residuals using dynamic weighting functions generated by an RNN is applied. Experimental results show that the RRSBR method significantly outperforms the SBR method.