Speaker
Description
Cochlear implants (CIs) aim to enable hearing in people with severe hearing loss by electrically stimulating the auditory nerve. Although highly successful, CIs fail to restore normal perception in many listening situations. We propose a deep learning framework to evaluate factors limiting CI-mediated perception in real-world auditory tasks. We optimized deep artificial neural network decoders to recognize words, localize sounds, and selectively attend to a cued talker in multi-talker situations using simulated auditory nerve representations as input. We trained decoders with either acoustically or electrically stimulated auditory nerve input to estimate upper bounds on hearing task performance given either normal hearing or CI-mediated peripheral representations. Once trained, we compared the models to humans with normal hearing and CI users by testing them on the same tasks. To investigate outcomes for different device manufacturers (Cochlear, MedEl, Advanced Bionics), we separately trained models on simulated nerve activity evoked by their respective sound-coding strategies.Models optimized with CI input exhibited impaired speech recognition, sound localization, and selective attention relative to normal hearing models, revealing limitations in the information conveyed by CI stimulation. Performance was similar across the different sound coding strategies. The best-performing human CI users approached the performance of the CI models when tested on the same task, consistent with the idea that the best-performing humans perform about as well as is possible given their CI. The ability to predict real-world behavioral outcomes for candidate prostheses opens the door to large-scale screening of new device strategies.