Explainable machine learning reveals the relationship between hearing thresholds and speech-in-noise recognition in listeners with normal audiograms

Aug 4, 2026

Upload by Finn

Some individuals experience difficulty listening in noise despite having normal audiograms. This study applies machine learning to assess how hearing thresholds predict speech-in-noise recognition in normal-hearing listeners. Performance of generalized additive modeling and four machine learning models was compared using archived thresholds from 0.25–16 kHz and speech recognition thresholds, with XGBoost showing the lowest mean absolute error. SHapley Additive explanations identified age and extended high-frequency thresholds as major contributors, supporting the relevance of hearing in extended high frequencies for predicting speech-in-noise recognition.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Abstract

Introduction

Motivation and prevalence of listening-in-noise difficulty

Clinical relevance of predicting speech recognition from audiograms

University of Texas Rio Grande Valley ScholarWorks @ UTRGV | School of Mathematical and Statistical
Sciences Faculty Publications and
Presentations | College of Sciences | | --- | --- | 10-12-2023 Explainable machine learning reveals the relationship between hearing thresholds and speech-in-noise recognition in listeners with normal audiograms Jithin Raj Balan Hansapani Rodrigo Udit Saxena Srikanta K. Mishra Follow this and additional works at: [https://scholarworks.utrgv.edu/mss_fac](https://scholarworks.utrgv.edu/mss_fac) Part of the Mathematics Commons, and the Speech Pathology and Audiology Commons ARTICLE ................................... Explainable machine learning reveals the relationship between hearing thresholds and speech-in-noise recognition in listeners with normal audiograms Jithin Raj Balan,1 Hansapani Rodrigo,2 Udit Saxena,3 and Srikanta K. Mishra1,a) 1Department of Speech, Language and Hearing Sciences, The University of Texas at Austin, Austin, Texas 78712, USA 2School of Mathematical and Statistical Sciences, The University of Texas Rio Grande Valley, Edinburg, Texas 78539, USA 3Department of Audiology and Speech-Language Pathology, Gujarat Medical Education and Research Society, Medical College and Hospital, Ahmedabad, 380060, India ABSTRACT: Some individuals complain of listening-in-noise difﬁculty despite having a normal audiogram. In this study, machine learning is applied to examine the extent to which hearing thresholds can predict speech-in-noise recognition among normal-hearing individuals. The speciﬁc goals were to (1) compare the performance of one standard (GAM, generalized additive model) and four machine learning models (ANN, artiﬁcial neural network; DNN, deep neural network; RF, random forest; XGBoost; eXtreme gradient boosting), and (2) examine the relative contribution of individual audiometric frequencies and demographic variables in predicting speech-in-noise recognition. Archival data included thresholds (0.25–16kHz) and speech recognition thresholds (SRTs) from listeners with clinically normal audiograms (n ¼ 764 participants or 1528 ears; age, 4–38years old) . Among the machine learning models, XGBoost performed signiﬁcantly better than other methods (mean absolute error; MAE ¼ 1.62dB) . ANN and RF yielded similar performances (MAE ¼ 1.68 and 1.67dB, respectively), whereas, surprisingly, DNN showed relatively poorer performance (MAE ¼ 1.94dB) . The MAE for GAM was 1.61dB. SHapley Additive exPlanations revealed that age, thresholds at 16kHz, 12.5kHz, etc., on the order of importance, contributed to SRT. These results suggest the importance of hearing in the extended high frequencies for predicting speech-in-noise recognition in listeners with normal audiograms. 2023 Acoustical Society of America. [https://doi.org/10.1121/10.0021303](https://doi.org/10.1121/10.0021303) (Received 20 March 2023; revised 4 August 2023; accepted 17 September 2023; published online 12 October 2023)[Editor: Christian Lorenzi] Pages: 2278–2288 I. INTRODUCTION Nearly 10% of patients who visit an audiology clinic complain of listening-in-noise difﬁculty despite having a normal audiogram (Alvord, 1983; Billings et al., 2018; Ferman et al., 1993; King, 1954; Middelweerd et al., 1990; Hind et al., 2011). Auditory and cognitive variables could contribute to such deﬁcits in adults and children (e.g., Dillon and Cameron, 2021; Pienkowski, 2017). The present study examines whether the most basic measure of hearing—thresholds—contributes to speech-in-noise recognition in listeners with clinically normal audiograms. A clinically normal audiogram was deﬁned as having hearing thresholds of 20dB hearing level (HL) or lower at octave frequencies from 0.25 to 8kHz. Measuring hearing thresholds is a fundamental aspect of hearing assessment. In fact, an audiogram may be the only metric for preparing a treatment plan for individuals with hearing loss in many under-resourced clinics.

FAQs

Why do some listeners with normal audiograms still have trouble understanding speech in noise?

Some individuals report listening-in-noise difficulty despite normal audiograms, suggesting that hearing thresholds may still carry predictive information not captured by the basic audiogram classification.

Which machine learning model performed best for predicting speech-in-noise recognition?

XGBoost performed significantly better than the other methods, achieving the lowest mean absolute error compared with ANN, RF, DNN, and GAM.

What did explainable AI indicate about which factors matter most?

SHapley Additive exPlanations showed that age and thresholds at extended high frequencies (e.g., 16 kHz and 12.5 kHz) contributed most to speech recognition thresholds, highlighting the importance of extended high-frequency hearing.

Aug 4, 2026

Upload by Finn