Conventional hearing aid (HA) algorithms, developed based on computational and non-differentiable auditory processing models, do not typically compensate for cochlear synaptopathy (CS). Traditional HAs apply fixed or rule-based gain adjustments within predefined frequency bands and compression ratios, instead of model-based fitting that numerically optimizes processing from neural representations of normal and impaired hearing. To compensate for combined CS and outer-hair-cell (OHC) loss, deep neural network (DNN)-based closed-loop HA systems are gaining traction. Here, we present several DNN-based HA algorithms that embed personalized, differentiable DNN-based auditory models (dCoNNear) inside a closed-loop system to train personalized HA algorithms compensating for OHC damage and/or CS. The HA algorithms were trained using backpropagation to minimize differences between hearing-impaired and normal-hearing auditory nerve (AN) responses. Performance was evaluated using speech and standard auditory stimuli. Results showed enhanced temporal-envelope (TENV) processing of modulated pure tones, particularly for CS, where sharpening of the TENV led to stronger AN onset responses. Transfer functions indicated that DNN-based HA algorithms applied adaptive level-dependent and frequency-specific gain aligned with OHC damage. The algorithms improved the normalized root mean square error of AN responses compared to NAL-NL2 for certain TIMIT phoneme categories. This evaluation offers insights into how machine-learning approaches outperform traditional HA strategies.
Wouters et al. (Sun,) studied this question.