Relationship between intelligibility, naturalness, and listening effort of source-separated speech

Various signal processing techniques have been proposed for separating a speech signal from an acoustic environment that includes other sound sources. One important measure of success of a source separation method is the intelligibility of the extracted speech signal, determined as the fraction of intended words correctly recognized. However, even for extracted signals with similar intelligibility, listeners may experience different reactions to the result. Speech samples extracted by different techniques may differ in the degree to which they sound pleasant or natural to a listener, and/or the degree of cognitive effort required to understand them. We present results of an experiment in which listeners were presented with context-free spondees recorded in a noisy environment and subsequently processed in various ways to enhance the speech and suppress the noise. Listeners transcribed the speech and also judged the “naturalness” and “listening effort” of the speech. Intentionally, stimulus intelligibility varied widely—as did naturalness and effort. We present and discuss the measured relationship among intelligibility, perceived naturalness, and reported listening effort.