Optimising sound coding for semi- and totally-implantable cochlear implants

A cochlear implant (CI) is a device that restores audition to the profoundly deaf by bypassing the impaired inner ear with a direct stimulation of the auditory nerve by means of patterns of electric current. In a CI, the sound is first captured by the microphone(s it is then processed with one or more enhancements stages such as noise reduction, and is finally converted to a sequence of electrical pulses by the stimulation strategy. All of these stages play a major role in the perception of sound by the implant user. Several technological advances in the last decades have made it possible for a CI user to attain relatively good speech understanding in quiet conditions. However, speech perception rapidly decreases in adverse listening situations such as in the presence of background noise. This work concerned all three stages mentioned above in the context of improving speech perception in noise for users of current semi-implantable CIs and future totally-implantable systems. Chapter 3 and 4 present NNSE, a noise reduction algorithm based on supervised machine-learning designed to investigate the relevance of these approaches to CIs, where real-time feasibility and generalisation performance to novel auditory scenes represent major challenges to overcome. The algorithm was successfully validated with normal-hearing (NH) subjects listening to CI vocoded simulations. The positive outcomes motivated further investigation with CI users. Here, significant speech perception improvements relative to the unprocessed condition were obtained with NNSE in stationary speech-shaped noise (SSN), multi-talker babble, and highly nonstationary ICRA noise when the algorithm was trained on novel segments of the same noise type and target speaker used in the listening experiments. Generalisation performance to novel speakers was tested by training NNSE on a speech dataset that did not include the target talker. In these conditions, smaller yet significant improvements were found for NNSE in SSN and ICRA noise, whereas in multi-talker babble the improvement was not statistically significant. Last, chapter 5 presents SPACE, a stimulation strategy that aims to improve the spectral representation of the sound by precompensating the stimulation pattern for the spread of excitation, one of the major information-transfer bottlenecks at the electrode-neuron interface. SPACE was evaluated with a group of cochlear implant listeners against their daily ACE programme in terms of preference rating and speech perception in SSN and four-talker babble noise. While no significant differences in preference rating were observed, results indicated that a statistically significant benefit in speech perception in noise could be attained with SPACE processing relative to ACE. Considering the acute nature of the listening experiments, the obtained results are encouraging and call for further investigation of this approach.