Segmental contribution to the intelligibility of noise-suppressed speech

Abstract

Many studies suggested that most speech enhancement algorithms do not improve the subjective intelligibility scores of noise-suppressed speech when presented to normal-hearing listeners. Nevertheless, the reason for lacking intelligibility improvement is still unclear. This study assessed the segmental contribution to the intelligibility of noise-suppressed speech. Mandarin sentences were corrupted by steady-state speech-spectrum shaped noise and multi-talker babble. Stimuli were synthesized by using noise-suppressed speech processed by three types of speech enhancement algorithms and a noise-replacement paradigm. The noise-replacement paradigm preserved selected speech segments (i.e., vowel-only and vowel-plus-consonant-onset), and replaced the rest with noise. Listening experiments were performed to collect the intelligibility scores from normal-hearing listeners. Experimental results showed that the selectively synthesized stimuli (i.e., vowel-only and vowel-plus-consonant-onset) from noise-suppressed speech may be more intelligible than those from noise-corrupted speech. However, this benefit was only observed when the speech signal was corrupted by speech-spectrum shaped noise but not by babble masker. This work suggested that the segmental distortion caused by speech enhancement algorithms may affect the information integration from noise-suppressed speech, and high-level top-down processing may account for the deficiency of the present speech enhancement algorithms for improving subjective intelligibility performance.