Theoretical analysis of parametric blind spatial subtraction array and its application to speech recognition performance prediction

In this paper, an improved parametric postfiltering is introduced in our previously proposed blind spatial subtraction array (BSSA), and its theoretical analysis of the amounts of musical noise and noise reduction is conducted via higher-order statistics. Compared with the conventional BSSA, it is clarified that parametric BSSA can improve speech recognition performance. Next, we propose an unsupervised speech-recognition-performance prediction metric based on higher-order statistics in BSSA. We successfully reveal that the noise and speech kurtosis can be used for predicting speech recognition performance without using any reference speech signals.