Speech Intelligibility Prediction for Cochlear Implant Recipients: An investigation of the Output Signal to Noise Ratio

Abstract

Cochlear implants (CI) can achieve excellent hearing outcomes for people with severe or profound hearing loss. However, outcomes vary significantly, and hearing in noise is challenging. Evaluation of new sound processing ideas with recipient testing is a lengthy process. A metric which reliably predicted CI speech intelligibility would allow more sound processing ideas and parameter sets to be evaluated. Prediction for normal hearing people has been extensively investigated but CIs have received little attention. In this thesis, it was hypothesised that an extended Output Signal to Noise Ratio (OSNR) metric would be a reliable predictor of individual CI speech intelligibility. A series of retrospective CI experiments found i. OSNR was sensitive to changes in a range of test parameters, including Input Signal to Noise Ratio (ISNR), presentation level, noise type, and processing algorithms. ii. OSNR predicted mean speech scores for a group of recipients at least as accurately as other tested metrics. Being calculated at the sound processor output OSNR predicted scores in scenarios where other metrics could not. iii. A novel method, combining OSNR and test scores from a reference condition, predicted psychometric functions for individual CI recipients relatively accurately in many other conditions. Small differences in performance between test and reference conditions were accurately predicted. OSNR was not accurate in the presence of aggressive spectral masking. The feasibility of combining OSNR and Output Speech Power as an accurate predictor in this scenario was demonstrated. The author is not aware of any published works which predict psychometric functions for individual CI recipients listening to speech processed by novel algorithms, using scores from a separate, reference condition. Overall, there was considerable support for the hypothesis that OSNR was an accurate predictor. OSNR shows promise as a metric with research and clinical applications.