Source-linked AI summary
Subjective Evaluation of DNN AND Auditory-Model-Based Hearing-Loss Compensation
Chuan Wen, Brent Nissens, Nele De Poortere, Morgan Thienpont, Matthias Inghels, Attila Fráter, Guy Torfs, Sarah Verhulst
TL;DR
OHC-related hearing loss affects cochlear tuning, and behavioral evidence for DNN-based compensation has been limited. This study personalizes an auditory-model-based hearing aid and tests it against unprocessed speech; intelligibility improved significantly, though the objective and behavioral evaluations were not strictly matched.
Problem
Behavioral benefits of auditory-model-based DNN hearing aids in human listeners remain insufficiently established despite objective promise.
Source: §1 · p. 1Method
The study personalizes a cochlear model to each listener’s audiogram, trains an OHC-targeting hearing aid, and compares unprocessed with processed speech in a Matrix test.
Source: §1 · p. 1, §2.4 · p. 3Results
OHC processing improved scores over unprocessed speech (Mdiff = 0.118, t(18) = 7.33, p < .001) across 19 participants.
Source: §3 · p. 4Takeaways & Limitations
The findings provide behavioral evidence of significant intelligibility benefit across the OHC-loss severity represented in the sample.
Source: §4 · p. 4The objective and behavioral evaluations were not strictly matched, and a behavioral comparison with NAL-NL2 remains future work.
Source: §3 · p. 4
Abstract
from arXiv · showhide
Outer-hair-cell (OHC) loss is a primary deficit of sensorineural hearing loss (SNHL), impairing cochlear amplification and frequency selectivity and thereby elevating hearing thresholds. Biophysically-inspired DNN-based hearing-aid (HA) algorithms have been proposed to compensate for OHC deficits and have shown clear benefits in objective speech intelligibility and quality metrics (e.g. HASPI, HASQI). However, comprehensive subjective validation of these benefits in human listeners is still missing. In this work, we present a subjective evaluation of a biophysically-inspired HA model targeting OHC deficits. The cochlear module of an auditory model was individualized based on each listener's pure-tone audiogram and integrated into a trainable system, which includes the personalized model and a normal-hearing reference model, and the resulting trained HA was evaluated using a Matrix test comparing intelligibility scores for unprocessed and HA-processed noisy speech. The results revealed a significant benefit of the HA model over the unprocessed condition in the range of +1 to +27%, providing behavioral confirmation of the efficacy of the HA model. This study closes the gap between objective and perceptual evidence for this new generation of DNN-based HA algorithms, paving the way for their integration into next-generation DNN-accelerated chips for hearables and hearing aids.
1. INTRODUCTION
OHC damage can elevate thresholds and broaden cochlear tuning, while conventional amplification does not compensate for broadened tuning. Although DNN-based auditory-model compensation has shown objective promise, its behavioral benefits in human listeners remain insufficiently validated.
1. INTRODUCTION
- OHC damage elevates hearing thresholds and broadens auditory filters, compromising frequency selectivity; conventional hearing aids do not compensate for broadened tuning.Conventional aids use frequency-dependent amplification and wide-dynamic-range compression (WDRC).Source: §1 · p. 1
- Auditory-model-based DNN hearing aids learn compensation from normal-hearing and hearing-impaired model responses, and have shown benefits on objective intelligibility and quality metrics.Training minimizes the discrepancy between the two pathways’ responses without predefined compensation constraints.Source: §1 · p. 1
- Prior subjective testing involved 13 listeners and a 61% unprocessed baseline, which the authors say could have produced ceiling effects that masked compensation gains.The authors identify the gap between objective and perceptual evidence as requiring further validation before real-world use.Source: §1 · p. 1
- The study evaluates an OHC-targeting dCoNNear hearing aid with a Matrix sentence-in-noise test, comparing intelligibility for unprocessed and processed speech.The authors report significant gains over the unprocessed baseline as behavioral evidence for the compensation strategy.Source: §1 · p. 1
2. METHOD
The method personalizes cochlear models using listeners’ audiograms, trains a hearing aid to reduce differences between hearing-impaired and normal-hearing auditory responses, and evaluates objective metrics and behavioral intelligibility. The behavioral test compares unprocessed and OHC-processed speech in noise.
2.1. Personalized auditory models
- Listeners’ audiograms personalize the cochlear model, and the resulting individualized hearing-impaired model is embedded in the dCoNNear hearing-aid training framework.The workflow begins with pure-tone audiometry and evaluates the trained aid behaviorally using a Matrix sentence-in-noise test.Source: §2 · p. 2, §2.1 · p. 2
2.2. Hearing-aid model training
- The hearing-aid training objective minimizes mismatch between hearing-impaired and normal-hearing simulated auditory responses while keeping both cochlear modules frozen.The loss averages mean absolute error across 10, 50, 100, and 200 ms temporal scales.Source: §2.2 · p. 2
2.3. Objective evaluation
- HASPIv2 and HASQI compare evaluated signals with clean references using an auditory-periphery model parameterized by the listener’s audiogram.The objective evaluation also compares OHC processing with NAL-NL2 for a fixed Slope35 5 audiogram and −3 dB SNR.Source: §2.3 · pp. 2–3
- OHC processing increased HASPIv2 compared with unprocessed and NAL-NL2 conditions, while yielding a lower HASQI score, indicating a trade-off between predicted intelligibility and measured signal quality.The comparison is reported for the evaluated processing conditions at −3 dB SNR.Source: Table 1 · p. 3, §2.3 · p. 3
2.4. Subjective evaluation
- The subjective test measures correctly identified words in Matrix sentences mixed with restaurant babble, comparing unprocessed and OHC-compensated conditions.Participants selected heard words from a 10×5 matrix; listeners were included only when baseline scores fell within the targeted 40–71% range.Source: §2.4 · p. 3
3. RESULTS
OHC processing improved speech intelligibility for most listeners, with a significant group-level benefit. Benefits occurred across the represented hearing-loss range, without a clear association between greater hearing loss and larger benefit.
3. RESULTS
- 17 of 19 listeners (89%) showed at least a +5% intelligibility benefit, and the group-level improvement was significant (Mdiff = 0.118, t(18) = 7.33, p < .001).The paired comparison covered 19 participants; Cohen’s dz = 1.68.Source: §3 · p. 4
- Benefits appeared across all three listener groups and the full EHF PTA range, without a clear association between greater hearing loss and larger benefit.The controlled baseline-performance range may partly explain the lack of association; Table 2 summarizes group-specific benefit ranges.Source: §3 · pp. 3–4, Figure 4 · p. 4, Table 2 · p. 4
- The subjective findings align with objective HASPIv2 predictions, but the evaluations used different audiograms and SNRs, and behavioral comparison with NAL-NL2 remains future work.The OHC model had a lower HASQI score than NAL-NL2, a quality dimension not assessed behaviorally here.Source: §3 · p. 4
4. CONCLUSION
The study reports a consistent, statistically significant intelligibility benefit from DNN-based OHC compensation over unprocessed speech, in agreement with its objective predictions. Benefits spanned the OHC-loss severity represented in the sample.
4. CONCLUSION
- DNN-based OHC compensation provided a consistent, statistically significant intelligibility benefit over the unprocessed baseline, with benefits across the OHC-loss severity represented in the sample.The subjective findings agreed with objective HASPIv2 and HASQI assessment in this study.Source: §4 · p. 4
- The study paired subjective evaluation of a biophysically-inspired DNN hearing-aid algorithm with objective assessment using HASPIv2 and HASQI.Source: §4 · p. 4
- Future behavioral evaluation should directly compare this approach with widely used strategies such as NAL-NL2.Source: §4 · p. 4