Assessing the perceptual contributions of level-dependent segments to sentence intelligibility

The present work assessed the contributions of high root-mean-square (RMS) level (H-level, containing primarily vowels) and middle-RMS-level (M-level, with mostly consonants and vowel-consonant transitions) segments to the intelligibility of noise-masked and noise-suppressed sentences. In experiment 1, noise-masked (by speech-spectrum shaped noise and 6-talker babble) Mandarin sentences were edited to preserve only H- or M-level segments, while replacing the non-target segments with silence. In experiment 2, Mandarin sentences were subjected to four commonly-used single-channel noise-suppression algorithms before generating H-level-only and M-level-only noise-suppressed sentences. To test the influence of an effective signal-to-noise ratio (SNR) on intelligibility, both experiments incorporated a condition in which the SNRs of H-level segments and M-level segments were matched. The processed sentences were presented to normal-hearing listeners to recognize. Experimental results showed that (1) H-level-only sentences carried more perceptual information than M-level-only sentences under both noise-masked and noise-suppressed conditions; and (2) this intelligibility advantage of H-level-only sentences over M-level-only sentences persisted even when effective SNR levels were matched, and it might be attributed to the perceptual advantage of vowels in speech intelligibility. In addition, the lesser distortion in H-level segments than in M-level segments following noise-suppression processing suggests that differential processing distortion might contribute to the H-level advantage observed.