Speech enhancement and intelligibility modeling in cochlear implants

UnrestrictedAccompanying cochlear implant (CI) performance improvement over years, CI speech recognition increasingly showed unique patterns and huge inter-subject performance difference that are different from that of normal hearing (NH) listeners. Previous literatures have not paid enough attention to such unique patterns, regardless of numerous evidences. To further improve speech perception with the next generation CI device, it is critical to integrate such unique patterns into the CI framework.; The thesis addressed speech enhancement and intelligibility modeling from systematic approaches. Speech enhancement utilized unique speech recognition patterns as feedback to modify input auditory signals to CI devices. Speech intelligibility modeling unified CI device settings and psychoacoustic responses from individual subjects to predict inter-subject performance difference.; Although speech recognition with NH listeners are highly robust even when acoustic information is lost or variant, such listening conditions are very challenging for CI users. A speech enhancement framework was proposed and realized to improve CI speech recognition in the context of telephone speech and different talker speech. The proposed speech enhancement framework significantly improved CI speech recognition; individual CI users showed substantially different recognition patterns in terms of the effect of speech enhancement, speech bandwidth and different talkers.; Speech intelligibility modeling was further studied to explore the observed inter-subject performance difference. An acoustic distance based intelligibility model was proposed to integrate variant CI device settings (e.g., frequency partition, electric stimulation rate, and acoustic-to-electric mapping) and psychoacoustic responses (e.g., electrode confusion patterns, intensity resolution, and dynamic range), aiming to capture complex and highly non-linear distortion of acoustic features through CI devices. Three phases of this work were involved. Firstly, a general speech intelligibility model was studied and validated under variant parametric effect in CI processing for averaged CI performance. Secondly, the effect of CI electrode confusion patterns was studied by smearing the spectral contrast of acoustic features. Thirdly, individual CI MAP and psychoacoustic response patterns were unified into the proposed model. Compared to the general speech intelligibility model, the customized model significantly boosted the linear prediction ability of inter-subject performance difference. The contributions of different psychoacoustic responses to inter-subject performance difference were further discussed