Experimental study on dereverberation and noise reduction for distant speech recognition

Room reverberation and environment noise pose significant challenges for both speech enhancement and distant speech recognition (DSR). However, the rapid progresses achieved recently in the two fields are somehow ignored by each other. Thus it is important to examine the benefit of the state-of-the-art front-end speech enhancement can provide for a typical DSR system. In this paper, we employ a recently-reported state-of-the-art dereverberation and noise reduction approach for DSR. This method can be evolved to several famous enhancement methods by choosing different parameters. We examine the correlations between seven commonly-used objective measures for enhancement evaluation and the speech recognition error rate of a typical DSR system. We show that, besides the substantial effectiveness of the enhancement approach for DSR, the perceptual evaluation of speech quality (PESQ) has the highest correlation with the speech recognition error rate over other object measures.