Modified ESTOI for improving speech intelligibility prediction

Abstract

Objective measures of speech intelligibility are commonly used as a practical alternative to expensive and laborious listening tests. The extended short-time objective intelligibility (ESTOI) measure has demonstrated high accuracy in predicting the intelligibility of speech corrupted by many types of degradation. In this paper we propose a modified version of ESTOI that is based on the glimpse model of speech perception in noise. Performance assessment against subjective data reveals that the modified ESTOI is equivalent to ESTOI on three data sets and slightly better than ESTOI on two data sets.