Comparing the perceptual contributions of cochlear-scaled entropy and speech level

Cochlear-scaled entropy (CSE) has been suggested to be a reliable predictor of speech intelligibility. Previous studies showed that speech segments with high root-mean-square (RMS) levels (H-levels) contained primarily vowels, which carry important information for speech recognition. The present work compared the contributions of high-CSE (H-entropy) and H-level segments to speech intelligibility. The natural speech was edited to generate two types of noise-replaced stimuli, which preserved the same percentages of largest CSE segments and highest RMS-level segments, and played to normal-hearing listeners in a recognition experiment. Experimental results showed that the nature of the noise-replaced stimulus, H-entropy and H-level, made a small difference in intelligibility performance. CSEs and RMS levels showed a moderately high correlation (r = 0.79), suggesting that many speech segments may have both large CSEs and high RMS levels, which might account partially for the small intelligibility difference b...