Biomimetic Cochlear Implant Coding: A Strategy Employing a Cochlear Model and Deep Neural Network

Objective: Current cochlear implant (CI) technology relies on simple filter banks that cannot fully emulate the natural processing of the cochlea. This limitation likely contributes to common user complaints: difficulty understanding speech in noise and poor music quality. To address this, this project will develop CI strategies that improve perception by more faithfully mimicking auditory nerve responses.

Method: To train a deep neural network (DNN) for cochlear implant stimulation, we first generated target normal hearing (NH) neurograms by applying audio signals to the CARFAC auditory model. These NH neurograms, representing auditory nerve responses, were used as input features for the DNN, which outputs 22 electrode stimulation currents. To evaluate the DNN’s output during training, these currents were fed into an electrical hearing (EH) model incorporating current spread, neural adaptation, and refractoriness, yielding a simulated CI neurogram. The DNN was then trained on open-source speech and noise materials, adjusting its weights to minimize the discrepancy (measured by a custom loss function) between the target NH neurograms and the resulting CI neurograms. This process can be seen in figure 1.

Results: An analysis employing the Structural Similarity Index (SSIM) and Mean Squared Error (MSE) indicated superior structural correspondence between NH neurograms and the CI neurograms produced via the CARFAC-DNN strategy, relative to those produced via the Nucleus ACE strategy.

Conclusions: The CARFAC-DNN strategy potentially offers a closer approximation of the natural auditory nerve response than established CI sound coding methods. Consequently, a perceptual evaluation within a sound booth environment is planned for CI recipients.