Noise reduction for cochlear implants

© 2014 Dr. Adam Anthony HersbachThis thesis is not authorised to be made available in the Baillieu Reading Room until 29 May 2019Cochlear implant (CI) users generally achieve acceptable speech understanding in quiet conditions, but have difficulty understanding speech in the presence of background noise. In this case, noise reduction processing can be utilised to help improve the situation, and solutions can be distinguished based on the number of microphones used to sample the acoustic environment. Single microphone solutions rely on the statistical properties of speech and noise while multi-microphone solutions can use the spatial characteristics of impinging sound to further separate speech from noise. It is the latter that forms the focus of the current research. A multi-microphone noise reduction algorithm was developed for a CI sound processor that attenuated sound from the rear while passing sound from in front of the listener. The algorithm used two microphones with small physical separation to generate two fixed directional patterns; one facing forward, the other towards the rear. By examining the front-to-back energy ratio, a signal-to-noise ratio (SNR) estimate was obtained, which was used to attenuate noise dominated frequency channels. The algorithm was evaluated acutely with CI listeners, primarily using an adaptive speech reception threshold (SRT) task, although sound quality and acceptable noise level were also studied. The acoustic environment used for evaluation in the laboratory included complex situations. These situations used various numbers of competing talkers or interfering speech weighted noise sources that changed spatial locations during the test. Reverberation was introduced and the algorithm was evaluated in a range of reverberant environments. Microphone sensitivity matching was investigated by introducing controlled levels of mismatch and measuring speech intelligibility performance. The evaluation revealed the algorithm was highly beneficial across a wide range of acoustic situations, outperforming a conventional generalised side-lobe canceller algorithm called Beam. The benefit varied with the spatial configuration of the competing noise and was greatest when the noise was located to the sides and rear of the listener. The benefit in reverberant conditions was maintained. Counter-intuitively, the benefits actually increased in the highest level of reverberation that was evaluated. Microphone mismatch had a detrimental effect on all multi-microphone algorithms that were evaluated, completely negating any multi-microphone benefit when the mismatch was 4 dB or greater. Finally, the algorithm was implemented in a wearable sound processor and CI users evaluated the algorithm outside the laboratory during their normal use of the device. Users were able to vote for their preferred listening program using their processor’s remote control device. The take-home evaluation consolidated the benefits measured acutely in the laboratory and provides critical guidance as to how the algorithm could be integrated into a commercial device