Using CARFAC-JAX, a fast, differentiable model of the human cochlea, to efficiently fit personalized hearing loss

Authors: Jason Mikiel-Hunter1, Nima Salimi1, Jessica Monaghan2, Alan Kan1, Jorg Buchholz 1

1Macquarie University
2National Acoustic Laboratories

Background: Recent machine-learning studies have highlighted the potential benefits of tuning amplification algorithms to the putative causes of an individual’s hearing loss. Typically artificial neural networks have been used to emulate the cochlea stage in these studies, massively reducing inference time. While this can reproduce the original cochlea model’s output well, it can also fix the feature space for the causes of hearing loss, impacting potentially the derived amplification algorithm’s effectiveness. Here, we demonstrate how the recent application of the Cascade of Asymmetric Resonators with Fast-Acting Compression model of the human cochlea in JAX (CARFAC-JAX) can overcome these issues, facilitating fast and efficient generation of personalized models of hearing impairment.

Methods: Pure-tone audiogram stimuli were presented to CARFAC-JAX (based on CARFACv2, Fig 1). Their detectability in the model’s steady-state output was measured relative to silence using a dprime measure (threshold equals 1). To permit spontaneous rates and suitable dynamic ranges in the “auditory nerve” output, the native CARFAC-JAX output (NAP, Fig 1) was transformed into a neurogram for “high spontaneous rate auditory nerve fibers” using a sigmoidal function. While gradient descent can be performed relative to any parameter in CARFAC-JAX, here we fitted individual hearing loss profiles as outer hair cell losses across the cochlea (OHC health), minimising 1-dprime for the 11 pure-tone audiogram frequencies in the associated loss function.

Results: Benchmarks for audio in CARFAC-JAX show a real-time factor (RTF) of ~0.02-0.11. This translates to a RTF of ~0.5-1.5 for differentiation relative to the OHC health parameter during each iteration of gradient descent. We further demonstrate that fitting personalized hearing losses in CARFAC-JAX is feasible (convergence within as little as ~3 minutes for 250 iterations), highlighting the necessary adaptations to achieve good fits.