Authors: Alejandro Ballesta Rosen¹; Jason Mikiel-Hunter²; Julian Maclaren¹; Jack Collins¹; Simon Carlile¹; Richard Lyon¹
¹Google Research; ²Macquarie University
Background: Conventional hearing aids rely on fixed, frequency-dependent amplification and compression, which often fail in complex environments like the “cocktail party” problem. To address the underlying encoding dysfunctions of hearing loss more comprehensively, we introduce the Differentiable Auditory Loop (DAL), an open-source framework for personalized hearing aid design and fitting.
Method: The DAL incorporates CARFAC, a differentiable model of human cochlear function (ported to JAX), to optimize a deep neural network that matches impaired auditory neural patterns with a normal-hearing reference. We utilized SEANet, a waveform-to-waveform fully convolutional UNet generator, fine-tuning it by comparing stabilized auditory images (SAIs)—2D representations of auditory nerve output—between normal and impaired CARFAC models. Through gradient descent, the SEANet model learns to simultaneously denoise the input and compensate for specific hearing loss characteristics.
Results: The DAL-optimized machine learning hearing aid was evaluated across five metrics measuring both audio interpretability and neural activity recovery. In all measures, the DAL framework outperformed a conventional master hearing aid (MHA) baseline, demonstrating superior restoration of auditory neural representations.
Conclusion: The DAL framework provides a novel and practical approach to hearing aid design that leverages machine learning for deeper personalization than the current state of the art. By integrating differentiable cochlear models directly into the optimization loop, this method moves beyond simple amplification. Next steps include hardware deployment to enable real-world clinical testing.
