Segmental contribution to predicting speech intelligibility in noisy conditions

It is necessary to identify speech segments carrying important information for speech intelligibility, particularly in noise. Earlier work based on a relative rootmean-square (RMS) level based segmentation suggested that middle-level (ranging from the overall RMS level to 10 dB below) segments contained more vowel-consonant boundaries wherein the spectral change was often most prominent, and perhaps most robust, in the presence of noise, and hence yielded improved performance of objective intelligibility modeling. Since the three levels (i.e., high-, middle- and low-levels) were defined empirically when proposed, the present work assessed how the boundaries of RMS-level based segmentation affected the performance of speech intelligibility prediction. When evaluated with speech recognition scores obtained with normal-hearing listeners and with a total of 72 noisedistorted and noise-suppressed conditions, it was shown that choosing 0 and - 10 dB to split middle-level led to maximized correlation in predicting the intelligibility of speech in noise.