Participant recruitment and inclusion criteria

We analyzed audiometric thresholds and word-recognition scores from 395 native English-speaking adults (18–80 years). These data were pooled from a series of studies (4/13/2017 – ongoing) that used consistent testing protocols and received approval by the Mass General Brigham Institutional Review Board [16,22–27]. All participants provided written informed consent at the time of their original enrollment. Cognitive screening was performed with the Montreal Cognitive Assessment (MoCA); only individuals with scores ≥ 26 were included. Additional inclusion criteria required good general health and no history of otologic or neurological disorders. At the time of testing, all participants underwent otoscopy and middle-ear screening using the Titan Suite from Interacoustics. Only ears with normal ear-canal volume, tympanic membrane mobility, and middle-ear pressure were included.

Audiometric threshold and word-recognition testing

Pure-tone air-conduction thresholds were obtained using an Interacoustics Equinox 4.0 audiometer (High Hz option). Standard thresholds were measured from 0.25 to 8 kHz (including 3 and 6 kHz) using TDH-45 headphones. To improve accuracy and reduce variability in the extended high-frequency (EHF) range, thresholds above 8 kHz (9, 10, 11.2, 12.5, 14, and 16 kHz) were measured with warble tones presented via HDA200 circumaural high-frequency headphones.

Speech materials were drawn from the Northwestern University Auditory Test No. 6 (NU-6; Auditec, Inc.), a clinical corpus of 50-item lists of phonemically balanced monosyllabic words [28,29], presented at suprathreshold levels (~40 dB SL; no less than 55 dB HL). Each word follows a consonant-vowel-consonant (CVC) structure and reflect the frequency of phoneme occurrence in American English, ensuring phonetic representativeness and minimizing lexical bias. The absence of semantic or contextual cues makes NU-6 well-suited for assessing suprathreshold speech recognition while limiting top-down influences. NU-6 recordings were sampled at 44.1 kHz; for modeling, stimuli were resampled to 100 kHz before auditory-periphery simulation.

To simulate challenging listening environments and stress peripheral temporal encoding, two signal degradations were applied to each NU-6 words: 1) time compression, which shortens overall duration and reduces cues at phonemic transitions; and 2) reverberation, which spreads temporally and spectrally by convolving each waveform with a synthetic room impulse response with a 0.3-s decay [30]. Specifically, all words were time-compressed to 65% of their original duration using a synchronous overlap-add (SOLA) algorithm and then reverberated with the same impulse response. The same modified NU-6 set was used in human testing and as input to the auditory-nerve model.

Simulation of auditory-nerve fiber responses to speech stimuli

To quantify how the three SR groups of ANFs contribute to spectro-temporal speech coding, we used a phenomenological model of the human auditory periphery [31]. This model captures key aspects of cochlear signal processing, including nonlinearities associated with basilar membrane filtering, synaptic transmission between inner hair cells and ANFs, and adaptation processes. It also incorporates stochastic properties of synaptic release and accounts partially for the diversity of ANFs by simulating distinct fiber populations with high (100 spikes/s), medium (5 spikes/s), and low (0.1 spikes/s) SRs. To approximate the variability in SRs observed across fibers, a fractional Gaussian noise generator was applied with randomized seeds on each stimulation trial, following prior model implementations [31,32]. All simulations used the human parameter set of the Zilany et al. model.

Responses were computed for 71 characteristic frequencies (CFs) logarithmically spaced between 125 Hz and 16 kHz and averaged across 200 repetitions. Outer and inner hair cell functions were set to normal (OHC = 1; IHC = 1) to isolate neural degeneration from pre-neural deficits. Synaptic adaptation was implemented via a cascade of power-law processes, approximated with infinite impulse response filters for computational efficiency [33].

Correlations between ANF responses and speech

Each time-compressed, reverberated waveform was resampled to 100 kHz, RMS-normalized to ensure consistent intensity scaling, converted to Pascals, and used as model input. The output was a neurogram: a two-dimensional representations of ANF firing rate as a function of post-stimulus time (bin width of 50 μs) and CF (71 CFs logarithmically spaced from 125 Hz and 16 kHz). Neurograms were generated separately for low-, medium-, and high-SR fibers.

To characterize the spectro-temporal features of the speech stimuli and to generate a basis for comparison with simulated neural responses, we applied a continuous wavelet transform (generalized Morse wavelet) to each waveform. This wavelet family offers flexible control over time-frequency tradeoffs and symmetry, making it well suited for analyzing speech dynamics. The real component was half-wave rectified to retain temporal fine structure and envelope fluctuations relevant to phase-locking while suppressing negative-going components that lack physiological interpretability. Wavelet coefficients and ANF responses were both sampled across the same 71 logarithmically spaced frequency channels (125 Hz to 16 kHz).

To quantify the fidelity with which different ANF types captured the spectral and temporal content of the stimulus, we calculated two-dimensional Pearson correlation coefficients between the wavelet spectrograms and the corresponding neurograms using the formula:

where W is two-dimensional matrix of the wavelet coefficient (in Pa) at time index m and frequency index n, v is the ANF firing rate (spikes/s) at the corresponding indices, and and are the matrix means.

To align responses to the acoustic input, spike latencies at each CF were adjusted for cochlear traveling-wave delay by selecting the lag that maximized the cross-correlation between the wavelet spectrogram and the neurogram in that CF channel.

Automatic word-recognition neural network

To estimate the perceptual consequences of CND, we trained Deep Neural Networks (DNNs) to classify the model-generated neurograms. Analyses were carried out in Python 3.9 with TensorFlow/Keras 2.16 20,21 [34,35]. Inputs were two-dimensional neurograms whose rows were the 71 CF channels and whose columns were 60,000 post-stimulus time bins. Neurograms were flattened to vectors of length (m· n), where m is the number of time bins and n is the number of CF channels, to match the input format of dense network layers.

Two fully connected architectures were evaluated to probe how decoder capacity influences sensitivity to SR-specific deafferentation. Neural Network 1 (NN1) comprised five dense layers (128 nodes each) followed by SoftMax output. Neural Network 2 (NN2) comprised sixteen dense layers (32 nodes each) to constrain capacity and encourage reliance on robust features that may better track human performance. Hidden layers used ReLU activation and dropout regularization (ρ = 0.03) followed each hidden layer to mitigate overfitting. The output layer had 25 units (the 25 NU-6 words) with SoftMax activation to yield a probability distribution over all lexical classes. Weights were initialized with Glorot uniform. Models were compiled with Adam (default settings), sparse categorical cross-entropy loss [34], and accuracy as the metric. Training was conducted using a fixed number of epochs with data shuffling prior to each epoch to reduce order-related bias. NN1 was trained for 10 epochs; NN2 for 500 epochs to ensure convergence given its greater capacity.

Training and evaluation corpus

The corpus comprised 25 NU-6 words time-compressed to 65% and reverberated with a 0.3-s decay. To promote generalization, 20 noisy variants per word were created by adding zero-mean Gaussian noise at SNRs from 11 to 30 dB in 1 dB steps. For testing, nine variants per word were generated at SNRs from −30–10 dB in 5 dB steps. Each waveform was resampled to 100 kHz, RMS-normalized, converted to Pascals, and presented to the auditory-periphery model at 90 dB SPL. A single presentation (no repetition averaging) was used to generate each neurogram. The dataset was split 69%/ 31% into training and held-out test set; the same split was used across conditions to enable paired comparisons.

Neurogram generation and CND manipulation

Three auditory-nerve population configurations represented different CND severities. Normal Innervation included 2 low-, 3 medium-, and 7 high-SR fibers per CF, approximating classic cat distributions [36]. High-SR only modeled selective loss of low- and medium-SR fibers while retaining 7 high-SR fibers per CF. Severe Neural Loss retained 3 high-SR fibers per CF (~75% ANF loss). OHC function was normal in all conditions. Performance was evaluated on the held-out test set using accuracy and confusion matrices to quantify how word recognition degrades as a function of neural survival pattern.

Statistical analysis

Analyses were performed in Python using open-source libraries, including Pingouin statistical package [37]. Two-sample independent t-tests compared two-dimensional correlation coefficients across fiber types. Pearson’s r assessed the relationship between age and word recognition scores with 65% time compression and reverberation (WRS_65%). To evaluate the relative contributions of age, standard audiometric frequencies (PTA_St), and extended high frequencies (PTA_EHF) to word recognition score, we fit the multiple linear regression:

WRS_65% ~ Age + PTA_St + PTA_EHF

Model outputs were examined to determine the statistical significance and relative contribution (e.g., standardized beta coefficients or relative importance metrics) of each predictor variable. Statistical significance was defined as p < 0.05. Where multiple pairwise tests were conducted, p-values were adjusted using Benjamini-Hochberg false discovery rate procedure. All preprocessing, modeling, and visualization were implemented in reproducible Python workflows.

Cochlear nerve degeneration and speech-in-noise deficits

A growing body of research supports CND as a contributor to speech-in-noise difficulties, even in individuals with clinically normal audiograms. To date, the only direct evidence linking CND to impaired speech intelligibility comes from histopathological analyses of human temporal bones, which have revealed widespread age-related CND [3,4,13,43] and have associated this neural degeneration with poorer word-recognition performance [43,44], even after accounting for audiometric thresholds [13]. All other supporting evidence remains indirect, relying either on associations between putative biomarkers of CND and suprathreshold perceptual deficits or on studies of clinical populations with conditions presumed to involve CND. In the latter, longitudinal or retrospective investigations have shown disproportionate speech comprehension difficulties in individuals with multiple instances of noise-induced temporary threshold shifts [45], sensorineural hearing loss [25], chronic conductive hearing loss [46], or sudden sensorineural hearing loss [47].

In the former group, several studies have provided physiological evidence consistent with CND-related speech-in-noise deficits that are not explained by elevated audiometric thresholds. As early as 2015, individual differences in envelope-following response (EFR) phase-locking strength and ABR wave I amplitude were correlated with self-reported listening difficulties in noise [48]. Later, electrocochleographic recordings in individuals with normal hearing sensitivity have revealed elevated SP/AP ratios consistent with CND, which were associated with reduced recognition of temporally degraded speech [26,49]. Subsequent work using stimuli optimized to recruit low- and medium-SR fibers showed that EFR magnitude correlated with performance on challenging word recognition tasks [23,50]. Comparable age-related effects have been documented by comparing EFR responses between younger and middle-age adults with normal audiograms, where reduced neural responses predicted greater speech-in-noise deficits and increased listening effort [51].

However, not all investigations reported consistent associations between neural biomarkers and perceptual deficits. In some cases, age-related reductions in ABR wave I growth functions, interpreted as indicative of CND, were not accompanied by measurable speech-in-noise impairments [52]. These inconsistencies may reflect a combination of inter-individual variability in cognitive and attentional factors, limitations of both speech-in-noise and electrophysiological assays to detect subtle impairments, and the complex interplay between peripheral and central auditory processing. Additionally, CND is highly correlated with outer hair cell loss [3,13], so controlling for audiometric threshold automatically reduces the apparent impact of neural loss on perception outcomes.

To better isolate the functional impact of CND, some studies incorporated broader test batteries that include cognitive, psychophysical, electrophysiological, and speech-in-noise measures using stimuli tailored to preferentially recruit low- and medium-SR ANFs. One such study found that neither hearing sensitivity, EFR metrics, nor executive-function measures accounted for the variability in spatial release from speech-on-speech masking [53]. Instead age, a variable highly correlated with CND, emerged as the primary predictor impacting speech-in-noise performance, even after adjusting for audiometric and cognitive factors.

Neuroimaging offers converging support. High-resolution magnetic resonance imaging (MRI) has revealed reduced cross-sectional area and altered microstructural properties of the cochlear nerve in older adults relative to younger controls, with reduced neural synchrony predicting weaker recognition of time-compressed and noise-degraded speech [54]. Magnetoencephalography combined with ABR measures further showed age-related reductions in medial geniculate body activity and cochlear neural onset responses linked to declines in speech-in-noise performance, highlighting combined effects of CND and central auditory pathway dysfunction on speech intelligibility in older adults [55].

Computational modeling prediction of speech-in-noise deficits

One computational modeling study estimated synapse counts between ANFs and inner hair cells by combining DPOAE and ABR wave I data. Lower predicted counts were associated with greater noise exposure, higher risk of tinnitus, advancing age, and poorer speech-in-noise performance [56]. However, most computational studies have predicted only modest or negligible perceptual consequences of synaptic loss, even under simulations involving dramatic reductions in ANF populations. Early approaches approximated CND by reducing auditory signal fidelity through stochastic undersampling [17,57], which produced greater impairment for brief stimuli (< 20 ms) and steeper threshold vs. signal duration functions in older adults with normal audiograms, but lacked physiological specificity and did not capture SR-specific roles in encoding. Another approach, grounded in signal detection theory, predicted minimal perceptual deficits under extreme deafferentation [18], while assuming equal information across fiber types and limiting performance by ANF variability alone.

More recent models incorporated richer biophysics and decoding stages. A modified Zilany et al. peripheral model [31] paired with a DNN trained to classify digits in noise showed significant intelligibility decline once simulated CND reached 50% [20]. A different auditory model [58] and a sensorineural hearing loss simulator suggested that up to 90% deafferentation might be required to elevate speech recognition thresholds in noise [19]. Another study simulated a large population of ANFs across SR classes and decoded spike patterns into speech features for resynthesis, finding threshold elevation in quiet and in noise with increasing deafferentation, with strongest effects beyond 90% loss [21].

Our work extends these prior efforts in three ways. First, we used ecologically demanding speech materials across multiple SNRs that stress temporal precision and dynamic range, two domains strongly dependent on low- and medium-SR fibers. Second, we simulated ANF responses with the Zilany et al. model [31] while explicitly varied SR-specific survival in line with histopathology. Third, we asked how CND-induced degradation of neurograms impacts decoding by comparing two fully connected NN architectures that differ in depth and width.

Decoder architecture proved pivotal. The shallow and wide network (NN1) quickly achieved high accuracy and was largely insensitive to fiber loss. The deeper, narrower network (NN2) showed CND-driven decrements that paralleled age-related declines in our behavioral cohort. This divergence suggests that over-parameterized decoders can mask biologically meaningful degradations by fitting idiosyncrasies in the degraded input, whereas constrained architectures better approximate human performance limits. We note that inputs were presented at 90 dB SPL to elicit robust suprathreshold responses across SR classes, a choice that likely pushes high-SR fibers toward saturation. This may understate their contribution at everyday levels and should temper direct quantitative comparisons.

Efferent control and implications for CND modeling

The current simulations omit efferent control. In vivo, the medial olivocochlear (MOC) pathway provides dynamic gain regulation at the cochlea, with fast and slow components acting over tens to hundreds of milliseconds [59–62]. These feedback loops can suppress masker-driven basilar-membrane motion, reduce saturation in high-SR fibers, and enhance envelope fidelity when backgrounds fluctuate. MOC drive is also state dependent: it increases with attention and task difficulty, varies across listeners, and can be weakened by aging or neuropathy. Excluding this pathway may therefore either overstate the impact of afferent loss when efferent support would have mitigated masking, or understate resilience in listeners who recruit strong MOC modulation during difficult tasks. Recent subcortical modeling work has begun to formalize these effects, showing that dynamic MOC gain control driven by cochlear-nucleus and inferior-colliculus inputs can account for forward masking and other temporal phenomena relevant to speech coding in noise [63,64]. It must be noted however, that minimally invasive assays of MOC reflex strength consistently show weaker MOC feedback inhibition in human subjects than in commonly used animal models such as guinea pig, cat or mouse [60,65,66], and morphological studies confirm that there is a much sparser MOC innervation in humans than in the well studied animal models [66].

Limitations and future directions

While our model accounts for a significant portion of the variability observed in human performance, it does not capture the full range of behavioral outcomes. Some of this gap may arise from the peripheral model itself, which captures many nonlinearities but may not reproduce all temporal dynamics of low-SR fibers, including adaptation and masking susceptibilities that are central under degraded conditions. Central compensation is also not modeled. We used monosyllabic words to minimize linguistic context and isolate peripheral coding, but everyday speech provides lexical and semantic redundancy that can mitigate peripheral deficits. Extending tests to sentences and fluctuating maskers will clarify how CND interacts with top-down context. Likewise, our DNN classifiers do not exploit the full range of strategies available to human listeners, which may limit direct numerical alignment.

Another caveat lies in the operational definition of SR classes. Our model adopts SR proportions from classic cat studies. SR distributions are species- and CF-dependent [9,36,67,68]. We also simulated complete loss of low- and medium-SR fibers in some conditions to bracket worst-case scenarios. These assumptions clarify the contributions of each fiber type but may not reflect the graded nature of CND in aging populations.