Developing a Novel Hearing Aid Framework Using Machine Learning and a Model of the Impaired Cochlea

Authors: Nima Salimi1, Jason Mikiel-Hunter1, Jessica Monaghan2, Arun Sebastian2, Jorg Buchholz1

1Macquarie University
2National Acoustic Laboratories

Background: Current assistive devices only approximately fit the hearing needs of individuals, offering amplification strategies that focus on restoring “first-order” audibility while struggling to improve speech intelligibility in challenging acoustic environments. This project aims to leverage advanced computational modelling of the inner ear, Cascade of Asymmetric Resonators with Fast-Acting Compression (CARFAC), and machine learning to design and fit a hearing aid for individual users.

Method: To explore the usage of Artificial Neural Network (ANN) as a hearing aid algorithm, we developed a training framework that minimizes the difference between outputs of a normal-hearing and a hearing-impaired CARFAC models (Fig.1). To generate the true responses (upper part of Fig. 1), the neural activity patterns (NAPs) of the normal CARFAC in response to the tone were first generated. Then, the smoothed naps (naps_sm_true) were used as the true responses during the training process. To generate the predicted responses, tones were first passed to the ANN, and then the ANN outputs were plugged to the impaired CARFAC (Hearing loss has been simulated by decreasing the OHC_health values of the CARFAC). Finally, the predicted smoothed naps (naps_sm_pred) were compared against the true counterparts (naps_sm_true) through the structural similarity index measure (SSIM ) to obtain the loss function. The final loss values used during the training were equal to 1-ssim (ssim: computed SSIM values).

Results and Conclusion: We could show that the model can converge using a single Dense layer and SSIM. This confirms that the proposed framework has the potential to be used for more complex acoustic signals, probably using more complicated neural network architectures. The proposed framework would also be tested with the state-of-the-art hearing aid algorithms such as Master Hearing Aid (MHA).