Modeling the impact of cochlear nerve degeneration on speech recognition performance.

Cochlear nerve degeneration (CND), including the loss of synapses between inner hair cells and auditory nerve fibers (ANFs), has emerged as a likely contributor to "hidden hearing loss", a condition in which listeners experience speech-in-noise difficulties that cannot be fully explained by audiometric thresholds. This form of primary neural de-afferentation preferentially affects low- and medium-spontaneous rate (SR) fibers, which are critical for encoding acoustic features such as amplitude modulations, especially under challenging listening conditions such as noisy backgrounds. Although CND is well established in animal models and post-mortem human studies, its perceptual consequences remain poorly understood due to the inability to directly assess synaptic integrity in living humans. Here, we combined behavioral testing, a phenomenological model of the auditory periphery, and deep neural network (DNN) decoding to quantify the perceptual impact of SR-specific fiber loss. Audiometric thresholds and word recognition scores for time-compressed, reverberant NU-6 words were obtained from 395 cognitively normal adults aged 18-80. To isolate the neural contribution to speech encoding, we simulated ANF activity under three CND profiles that varied the survival of SR classes, transformed responses into time-frequency neurograms, and decoded them with two DNN architectures trained on word classification. Both networks learned the task, but only the deeper, more constrained model produced recognition scores consistent with human performance and showed sensitivity to CND, with recognition declining as low- and medium-SR fibers were removed. These findings provide a mechanistic link between SR-specific synaptopathy and speech-in-noise difficulties and establish a computational framework for evaluating the perceptual impact of hidden hearing loss.