Assessing the effect of noise-reduction to the intelligibility of low-pass filtered speech

Given the fact that most hearing-impaired listeners have low-frequency residual hearing, the present work assessed the effect of applying commonly-used singlechannel noise-reduction (NR) algorithms to improve the intelligibility of low-pass filtered speech, which simulates the effect of understanding speech with low-frequency residual hearing of hearing-impaired patients. In addition, this study was performed with Mandarin speech, which is characterized by its significant contribution of information present in (low-frequency dominated) vowels to speech intelligibility. Mandarin sentences were corrupted by steady-state speech-shaped noise and processed by four types (i.e., subspace, statistical-modeling, spectral-subtractive, and Wiener-filtering) of single-channel NR algorithms. The processed sentences were played to normal-hearing listeners for recognition. Experimental results showed that existing single-channel NR algorithms were unable to improve the intelligibility of low-pass filtered Mandarin sentences. Wiener-filtering had the least negative influence to the intelligibility of low-pass filtered speech among the four types of single-channel NR algorithms examined.