Which phonetic contrasts recover information lost to aggregate speech scoring in auditory nerve disorders: a computational framework

AO_SCPLOWBSTRACTC_SCPLOWAuditory nerve disorders, including auditory neuropathy spectrum disorders (ANSD), show distorted patterns of auditory nerve activity despite preserved spectral analysis of sound in the cochlea, producing impaired speech recognition despite normal auditory sensitivity. Distinct pathophysiological mechanisms affecting auditory-nerve activity have been identified in animal models, but current clinical speech tests were not designed to distinguish among them, because they average performance across phoneme categories, collapsing mechanism-specific confusion patterns into a single intelligibility score. Using computational modeling of auditory nerve responses, we simulated four candidate mechanisms and showed that their encoding disruptions cascade into systematic, phoneme-specific recognition failures: brief consonants were severely degraded while sustained vowels were preserved, and each mechanism produced a distinct confusion pattern. Models trained on ANSD-degraded signals developed compensation strategies that generalized to healthy signals, while the reverse did not, and noise training that benefited healthy models harmed ANSD models, mirroring real-world listening difficulty. Because simulation fixes the generating mechanism by construction, we could then measure directly how much mechanism-specific information each behavioral representation preserves. Aggregate word scoring discarded most of the information retained in the full confusion matrix, yet a handful of targeted phonetic contrasts recovered the mechanism. Simulation uniquely enables this analysis because the underlying mechanism is known, a property currently unavailable in patient datasets. Rather than offering direct diagnosis from behavior alone, the framework quantifies what aggregate scoring destroys and identifies the contrasts that best separate mechanisms, a principled route toward phonetically efficient tests that recover mechanistic information current scoring discards.