Development of dereverberation algorithms for improved speech intelligibility by cochlear implant users

Although reverberation does not significantly affect speech quality and intelligibility for normal hearing (NH) listeners, it severely degrades word identification performance in hearing impaired (e.g., cochlear implant (CI) users) as well as elderly people.;Several algorithms have previously been proposed to overcome the destructive effects of reverberation on speech quality. However, less work has addressed the reverberation effects on speech intelligibility in CI users. Word recognition performance of cochlear implantees in reverberant conditions is studied and new reverberation suppression strategies are proposed in order to improve intelligibility. Effectiveness of the proposed techniques is first evaluated in ideal scenarios, where access to an anechoic signal is assumed, to determine an upper-bound on performance. Formal listening experiments indicate large speech intelligibility gains in low to highly reverberant environments which motivates development of non-ideal suppression methods. We propose two novel channel selection criteria to blindly (non-ideally) suppress the reverberation in CI devices. These techniques lead to significant improvements for CI subjects in word identification performance in reverberation and can easily be integrated into current CI devices.;In addition to reverberation, the presence of background noise can also degrade word identification scores of CI users. Noise masks speech in a different manner compared to reverberation, and when added to reverberation, further elevates the confusion for listeners compared to reverberation/noise alone conditions. The combined effects of noise and reverberation are quite detrimental to speech intelligibility especially for CI users with only a limited number of stimulating electrodes in their devices. In this dissertation, the performance of CI listeners is evaluated in more realistic environments where reverberation and noise co-exist.;The destructive effects of reverberation and noise on speech intelligibility is assessed in order to quantify which of the two affects speech intelligibility the most. Moreover, noisy speech processed by two alternative ideal channel selection techniques is studied and compared with the anechoic quiet speech processed by ACE/CIS strategies, in terms of quality and intelligibility. Results indicate no significant difference in quality and intelligibility of the noisy speech processed by these alternative techniques and that of the ACE/CIS processed anechoic clean speech.