Title: The Differentiable Auditory Loop (DAL): An ML Framework for Hyper-Personalized Hearing Aids

Abstract:

Traditional hearing aids typically depend on static amplification and compression strategies tied to specific frequencies to compensate for diminished sensitivity. However, this approach frequently falls short in delivering adequate auditory support within complex acoustic settings, particularly during the "cocktail party" scenario involving multiple concurrent speakers. To better address the root encoding deficits associated with hearing loss, we present the Differentiable Auditory Loop (DAL), an open-source framework designed for the customization and fitting of hearing aids.

In our initial deployment of DAL, we utilized CARFAC—a differentiable model replicating human cochlear mechanics—adapted for JAX. This allowed us to train a deep neural network to align auditory neural activity patterns from impaired ears with those of a normal-hearing reference. To achieve the precise spectro-temporal signal processing necessary for effective hearing assistance, we employed SEANet, a fully convolutional UNet generator capable of waveform-to-waveform transformation.

The network’s fine-tuning process involves comparing the outputs of two CARFAC models: one calibrated to normal hearing and another adjusted to reflect an individual’s specific hearing impairment. This comparison utilizes loss functions based on the neural activity pattern (NAP) outputs from each CARFAC model, as well as stabilized auditory images (SAIs). SAIs offer a two-dimensional representation that encapsulates the phase-insensitive temporal structure found in auditory nerve outputs. Through gradient descent, the SEANet model acquires the ability to reduce background noise and correct for hearing loss as simulated by the impaired CARFAC model.

Evaluations across various neural-representation and signal-fidelity metrics indicate that the DAL-optimized SEANet model surpasses existing master hearing aid (MHA) baselines. The DAL framework offers a viable route toward machine-learning-driven, model-based personalization of hearing aid signal processing. Future efforts will focus on hardware integration to facilitate real-world clinical trials.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC