A Model for Predicting Intelligibility of Binaurally Perceived Speech

Predicting and modeling intelligibility of monaurally or binaurally presented speech is difficult because it depends primarily on the accuracy and interdependency of frequency, time, and spatial information arriving at the listener. Despite these complex relationships, a new pragmatic model is suggested for speech mixed with broadband noise. A form of the logistic regression function is used to characterize human performance data. The regression of these signal properties onto empirical speech recognition performance data estimates the relationship of these properties to speech recognition. This concept is illustrated by the modeling of human performance on Central Institute for the Deaf W-22 speech items presented monaurally and binaurally in both reverberant and non-reverberant conditions at different signal-to-noise ratios. Although the implementation of the present model is limited to the data considered, it is expected that other data can be modeled after the procedure outlined in this report. The model described is the first step in developing an objective binaural measure for predicting speech perception in noisy environments. The original document contains color images.