Utilizing cross-channel features for reverberation mitigation in cochlear implants

Cochlear implant (CI) users experience difficulties in understanding speech in reverberant environments as both active speech and non-informative speech segments of the reverberant signal are incorporated into the CI stimulus pattern. The non-informative speech segments are due to late reverberant signal reflections and are referred to here as overlap-masking segments. Desmond et al. (JASA, 2014) showed that removal of all overlap-masking segments from the CI stimulus pattern significantly improved the intelligibility of reverberant speech in CI users. Desmond (2014, Ph.D. dissertation) proposed a reverberation mitigation strategy based on training a machine learning classifier to identify and delete overlap-masking segments in CI pulse trains. The mitigation strategy employed within-channel features to mitigate overlap-masking effects in a given electrode channel. However, since adjacent electrode channels tend to enter the overlap-masking state concurrently, additional information about the state of oth...