Model-Guided Design of Low-Context Speech Probes for Cochlear Synaptopathy

Ahsan J. Cheema; David Meng; Jorge Mejia; Sanna Hou; Sunil Puria

Source-linked AI summary

Model-Guided Design of Low-Context Speech Probes for Cochlear Synaptopathy

Ahsan J. Cheema, David Meng, Jorge Mejia, Sanna Hou, Sunil Puria

arXiv:2609.36272v1Sep 28, 2026eess.AS

TL;DR

Cochlear neural degeneration can impair suprathreshold coding without elevating audiometric thresholds, complicating diagnosis when hair cell loss coexists. The paper compares temporal and noise-based probes with an auditory-nerve MI framework using VCV syllables, then tests model-selected conditions behaviorally. Combined compression and AM noise yielded the largest modeled information loss and was the only tested condition that distinguished the listener groups.

Problem
CND can impair suprathreshold coding without elevating thresholds, and its effects can be difficult to distinguish from hair cell dysfunction when both reduce neural output.
Source: §1 · p. 1
Method
The study modeled MI-based information loss for temporal and noise-based probes using VCV syllables across seven CND profiles, then tested model-selected conditions in a consonant-identification study.
Source: §1 · p. 1, §2.1 · p. 2, §2.2 · p. 2
Results
40% compression in AM noise produced the largest modeled ∆AUC at approximately 14.5 bits·log(Hz), and compressed speech in AM noise was the only tested condition with a significant accuracy difference between groups.
Source: §3.2 · p. 3, §3.3 · p. 3
Takeaways & Limitations
The findings provide partial support for using modeled information loss to guide stimulus design, while motivating further evaluation of combined temporal and noise-based probes for CND detection.
Source: §4.2 · p. 4, §5 · p. 4
The behavioral group difference cannot yet be attributed specifically to CND, and the full model-based stimulus ranking was not tested behaviorally.
Source: §4.2 · p. 4

Abstract

from arXiv · show

Cochlear neural degeneration (CND) can impair suprathreshold coding without elevating pure-tone thresholds, complicating its diagnosis when it coexists with hair cell loss. We present a unified comparison of temporal and noise-based probes for CND detection using low-context vowel-consonant-vowel (VCV) syllables to reduce linguistic and contextual cues. Using a phenomenological auditory nerve model, we simulated responses to 21 VCV tokens under time compression, reverberation, and speech-in-noise conditions across presentation levels and seven CND profiles. We computed mutual information (MI) between inner hair cell potentials and auditory nerve neurograms and quantified information loss relative to a normal-hearing baseline. Time compression and amplitude-modulated (AM) noise produced the largest modeled information losses. We then evaluated these stimuli in a consonant-identification study involving 36 listeners with normal audiograms, 12 of whom reported difficulty understanding speech in noise. Neither 40 percent time compression in quiet nor AM noise alone distinguished listeners with and without these difficulties. However, compressed speech presented in AM noise separated the two groups. This partial agreement between model predictions and behavior supports our MI-based stimulus design framework and motivates further evaluation of combined temporal and noise-based probes for CND detection.

1. INTRODUCTION

Cochlear neural degeneration can impair suprathreshold coding despite normal audiometric thresholds, and its effects are difficult to distinguish from hair cell dysfunction. This study compares temporal and noise-based probes using low-context VCV syllables and tests whether model-selected conditions distinguish listeners with and without self-reported speech-in-noise difficulty.

Research problem

CND can impair suprathreshold coding without elevating thresholds, while hair cell dysfunction and neural loss can both reduce neural output and complicate psychophysical diagnosis.This challenge is especially relevant when CND coexists with hair cell loss.Source: §1 · p. 1

Motivation

Time compression and reverberation alter speech timing, while noise probes target the hypothesized greater role of CND-vulnerable low- and medium-SR fibers under masking.A prior MI study using CVC words found the largest modeled information loss with time compression.Source: §1 · p. 1

Study aim

The study extends MI analysis to VCV nonsense syllables and compares temporal and noise-based probes across presentation levels and seven CND profiles.It tests the two model-ranked conditions individually and together in listeners with normal audiograms, with or without self-reported speech-in-noise difficulty.Source: §1 · p. 1

2. METHODS

The study compares temporal and noise-based VCV probes using auditory-nerve simulations and evaluates model-selected conditions in a consonant-identification task. Its methods span stimulus construction, modeled CND profiles, MI-based information-loss estimates, and behavioral testing.

2.1. Speech Corpus and Probe Conditions

The corpus comprised 21 VCV nonsense tokens, processed with compression, reverberation, stationary noise, or AM noise across four simulation levels.AM noise used 4 Hz sinusoidal modulation at 100% depth; simulations used 0 dB SNR.Source: §2.1 · p. 2

2.2. Auditory Nerve Model and CND Profiles

The auditory nerve model simulated IHC potentials and ANF neurograms across seven fiber profiles, including intact and six neural-loss populations.Profiles varied low-, medium-, and high-spontaneous-rate fiber counts, reflecting reported preferential vulnerability of low- and medium-rate fibers.Source: §2.2 · p. 2

2.3. Mutual Information Analysis

MI between IHC potentials and ANF neurograms was estimated at each characteristic frequency, then information loss was calculated against a normal-hearing reference.The MI estimate used a joint histogram with 1024 bins.Source: §2.3 · p. 2

2.4. Behavioural Validation Study

Thirty-six adults with clinically normal audiograms were divided into control and self-reported hearing-difficulty groups; noise was presented at each listener’s SRT50.Participants identified the medial consonant in a closed-set task, and accuracy and reaction time were analyzed with mixed-effects models.Source: §2.4 · pp. 2–3
The behavioral study tested unprocessed and 40%-compressed VCV tokens in quiet or AM noise, with noise conditions selected from model comparisons.The figure caption identifies 40% compression as the largest-loss temporal probe across seven CND profiles.Source: §2.4 · p. 2, Figure 1 · p. 3

3. RESULTS

The combined temporal-and-noise condition produced the largest modeled information loss and was the only tested condition that distinguished the listener groups. Time compression also led the temporal probes in quiet, while behavioral evidence remained limited to the tested conditions and group comparison.

3.1. Temporal Probes in Quiet

Approximately 12.5 bits·log(Hz): 40% compression produced the largest modeled ∆AUC among temporal probes at the most severe CND profile.This exceeded 65% compression at approximately 10.5 bits·log(Hz); reverberation reduced the values, and reverberation alone produced approximately 7.8 bits·log(Hz).Source: §3.1 · p. 3

3.2. Noise Type and Time Compression

Approximately 14.5 bits·log(Hz): 40% compression in AM noise produced the largest modeled ∆AUC, exceeding 65% compression in AM noise at approximately 12 bits·log(Hz).AM noise alone reached approximately 10 bits·log(Hz), compared with approximately 8.5 bits·log(Hz) for stationary noise at the most severe profile; increasing level or reducing SNR further produced no appreciable increase.Source: §3.2 · p. 3

3.3. Behavioural Validation

42.3% versus 47.9% accuracy: compressed speech in AM noise was the only tested condition with significantly lower accuracy in the SRHD group than Controls.The groups did not differ significantly for unprocessed speech in AM noise; reaction-time differences involving Group were also nonsignificant.Source: §3.3 · p. 3

4. DISCUSSION

The model’s largest-loss condition also distinguished the listener groups behaviorally, but this agreement was partial and does not establish CND as the cause. Task ceiling and listeners’ ability to use remaining cues further qualify how modeled information loss maps to identification.

4.1. Effects of Time Compression and AM Noise

Compression shortens speech cues and event intervals, while AM noise alternates masking; these demands may interact with modeled synaptic adaptation and recovery.Above 65 dB SPL or below 0 dB SNR, noise conditions produced little additional modeled ∆AUC within the tested conditions.Source: §4.1 · p. 3

4.2. Behavioural Support for the MI Framework

40% compression in AM noise produced the largest modeled ∆AUC in the study.Source: Figure 2 · p. 4
Only compressed speech in AM noise distinguished Control and SRHD listeners (planned comparison, p = .041), partially supporting the model-guided stimulus choice.Source: §4.2 · pp. 3–4, Figure 3 · p. 4
The behavioral difference arose at individual SRT50 values about 16 dB below the primary simulation SNR, and the complete stimulus ranking was not behaviorally tested.Similar audiometric thresholds and nonsense syllables limited some confounds, but central contributions remained and CND was not established as the cause.Source: §4.2 · p. 4

4.3. Information Loss and Performance in Quiet

Near-ceiling closed-set accuracy may explain quiet-condition divergence: modeled information loss does not specify task information needs, and perceptual adaptation is absent from the peripheral model.Both groups scored 100% for unprocessed speech and above 97% after compression in quiet.Source: §4.3 · p. 4

5. CONCLUSION AND FUTURE DIRECTIONS

The MI-based auditory nerve model identified compressed VCV syllables in AM noise as the largest modeled information-loss condition, and this was the only tested condition that significantly distinguished the listener groups. The authors call for further physiological testing to assess the probe’s relation to cochlear neural dysfunction.

40%-compressed VCV syllables in AM noise produced the largest modeled information loss and were the only tested condition to distinguish listeners with and without self-reported hearing difficulty.Source: §5 · pp. 4–5
The behavioral difference cannot yet be attributed specifically to CND; establishing probe sensitivity and specificity remains future work.Planned physiological measures include EFRs, ABR wave I, and middle-ear muscle reflexes.Source: §5 · pp. 4–5

6. GENERATIVE AI DISCLOSURE

The authors disclose that generative AI assisted only with manuscript editing and polishing; scientific work and intellectual contributions remained theirs.

Generative AI assisted only with editing and polishing; the authors take responsibility for the scientific content and intellectual contributions.Source: §6 · p. 5