Developing novel electrical stimulation strategies for cochlear implant users based on a model of the healthy human cochlea

Authors: Maryam Hosseinir1, Tim Brochier2, Zachary Smith2

1Macquarie University

Background: Many cochlear implant (CI) recipients struggle to engage in conversations in noisy environments and report dissatisfaction with music quality. Current CI systems use simplified filter banks, lacking full simulation of natural cochlear processing which may contribute to the limitations of the device. In this project, we aim to enhance CI strategies to better replicate the auditory nerve responses in the healthy human cochlea using a sophisticated auditory model, the Cascade of asymmetric resonators with fast-acting compression (CARFAC).

Method: A deep neural network (DNN) was trained to produce 22 electrode stimulation envelopes from audio inputs. These envelopes were then passed through a convolutional neural network (CNN). The CNN was trained to mimic an electrical hearing model that takes current spread, neural adaptation, and refractoriness into account. The weights of the CNN were then fixed, and the DNN-CNN was trained to minimize the difference between the electrical and acoustic hearing models.

Results: We trained all the models on sentences from the TIMIT database. The envelopes used as targets for the DNN and as input to the CNN were obtained by passing sentences from TIMIT dataset through advanced combination encoder (ACE) algorithm which is currently used in Cochlear Pty Ltd CIs. The output of CARFAC was used as a target for matching neurograms to normal hearing neural activity. We show that our DNN-CNN architecture achieves a higher similarity to CARFAC outputs compared to Neurograms obtained from ACE, i.e., a higher structural similarity index and a lower mean squared error (Figure 1).

Conclusion: Our method of producing electrode stimulation envelopes results in neurograms that resemble the simulated neural activity of a healthy cochlea, more than those produced by ACE. The perceptual significance of this method compared to ACE is being investigated.