A hearing impairment simulation method using audiogram-based approximation of auditory charatecteristics

Nozomi Jinbo, Shinnosuke Takamichi, Tomoki Toda, Graham Neubig, Sakriani Sakti, Satoshi NakamuraInterspeech 2014 · OpenReview

Abstract

This paper presents a personalized hearing impairment simulation that approximates complex auditory filter parameters using only a standard, easily measurable audiogram. By employing linear regression to estimate filter shapes and combining this with a gain reduction process, the proposed system replicates both the volume loss and the loss of frequency clarity (spectral smearing) characteristic of sensorineural hearing impairment. Experimental results demonstrate that the simulation effectively matches the word recognition rates and phoneme confusion patterns of specific hearing-impaired individuals when tuned with an interpolation weight.

Introduction

Hearing loss is often misunderstood as a simple reduction in volume. For individuals with sensorineural hearing impairment, the primary challenge is frequently a loss of "frequency selectivity," which causes spectral smearing. This makes speech sound muffled or distorted even when amplified. While high-fidelity simulations of these effects exist, they typically require time-consuming clinical measurements of auditory filters. This research addresses the need for a practical, personalized simulation that can be quickly configured using a standard audiogram.

Methodology

The simulation framework consists of two primary digital signal processing stages: a gain reduction process and a smearing process.

Overview of the simulation framework, detailing the Gain Reduction and Smearing processes.

Gain Reduction Process

This stage reduces the power of the input waveform to simulate the threshold of hearing. It uses a piece-wise linear function to map input sound pressure levels to reduced levels, effectively performing the inverse operation of a hearing aid.

Smearing Process

The core innovation lies in modeling the auditory system using rounded-exponential (roex) filters whose parameters are estimated via linear regression. Instead of direct measurement, the filter parameters ($p_u$ and $p_l$) are predicted based on the decibel values from an individual's audiogram. To account for the gap between raw physical simulation and the cognitive compensation strategies of hearing-impaired persons, an interpolation weight (alpha) is introduced. This allows users to blend between a normal hearing profile and the predicted impaired profile.

Results and Discussion

The simulation was evaluated using familiarity-controlled word lists (FW07). The study confirmed that simulating both gain reduction (G) and smearing (S) is essential; the combined "GS" effect results in a significantly more accurate representation of the difficulties faced by hearing-impaired individuals than either effect alone.

Word correct rates for severe and moderate hearing-impaired profiles compared to the simulation (GS) across different interpolation weights (alpha).

As shown in the evaluation, by adjusting the interpolation weight alpha to between 0.4 and 0.8, the word recognition rates for normal-hearing listeners using the simulation closely matched the actual performance of the target hearing-impaired individuals. Furthermore, the simulation successfully replicated specific phoneme confusion patterns, such as the confusion between the vowels /i/ and /u/ in severe cases. The impact of the simulation was notably more severe on unfamiliar words, highlighting the reliance of hearing-impaired individuals on linguistic context.

Conclusion

The proposed method provides a viable "shortcut" for creating personalized hearing impairment simulations using standard clinical data. While the current model uses fixed parameters for uncomfortable hearing levels and was tested on a limited sample, it offers a powerful tool for education and clinical empathy. Future work will focus on expanding the dataset to include more diverse hearing loss profiles and refining the gain reduction model to account for individual recruitment effects.

Summary updated

BibTeX
@inproceedings{jinbo2014hearing,
  author = {Nozomi Jinbo and Shinnosuke Takamichi and Tomoki Toda and Graham Neubig and Sakriani Sakti and Satoshi Nakamura},
  title = {{A hearing impairment simulation method using audiogram-based approximation of auditory charatecteristics}},
  booktitle = {Interspeech},
  year = {2014},
  url = {https://openreview.net/forum?id=p_hf6NF7Ey},
}
Figure