Classification of Age-Related Hearing Loss using Conversational Speech Biomarkers
(2025)Department of Automatic Control
- Abstract
- Hearing loss is increasing with an ageing population. There is thus a need for more convenient diagnosis methodology. This thesis investigates whether it is possible to classify age-related hearing loss using speech biomarkers and Automatic Speech Recognition (ASR) or not. The methodology was a binary classification problem based on analysing conversational data between a participant with normal hearing and a participant with age-related hearing loss. These conversations took place in nine different conditions with varying background noise, and levels of hearing aid support for the participant with hearing loss. The ASR model used was Whisper by OpenAi™. The extracted features from the data included frequency characteristics, spectral... (More)
- Hearing loss is increasing with an ageing population. There is thus a need for more convenient diagnosis methodology. This thesis investigates whether it is possible to classify age-related hearing loss using speech biomarkers and Automatic Speech Recognition (ASR) or not. The methodology was a binary classification problem based on analysing conversational data between a participant with normal hearing and a participant with age-related hearing loss. These conversations took place in nine different conditions with varying background noise, and levels of hearing aid support for the participant with hearing loss. The ASR model used was Whisper by OpenAi™. The extracted features from the data included frequency characteristics, spectral attributes, power measures, pauses, floor-transfer offsets, Mel Frequency Cepstral Coefficients, and principal components derived from the first and final encoder block outputs of the Whisper model. The impact of the conditions and the features on the classification accuracies were analysed in conjunction to evaluate an optimal future method to detect hearing loss from speech data. The results suggest that it is possible to classify age related hearing loss with speech data, the top accuracy achieved was 78.4%. The background noise and the level of hearing aid support affected the outcome of the classification accuracy, as well as the selected features. The most influential features were related to power, frequency, and the Whisper encoder block outputs. This thesis does not incorporate control group conversations between participants with normal hearing, and the data set is limited to 24 participant pairs. To achieve more reliable and generalisable results, the inclusion of a control group and a larger data set is essential. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9183947
- author
- Bergström, Sanna and Ekre, Fremja
- supervisor
- organization
- alternative title
- An Implementation of Feature Engineering and ASR in a Machine Learning Pipeline
- year
- 2025
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6266
- other publication id
- 0280-5316
- language
- English
- id
- 9183947
- date added to LUP
- 2025-02-04 14:06:32
- date last changed
- 2025-02-04 14:06:32
@misc{9183947,
abstract = {{Hearing loss is increasing with an ageing population. There is thus a need for more convenient diagnosis methodology. This thesis investigates whether it is possible to classify age-related hearing loss using speech biomarkers and Automatic Speech Recognition (ASR) or not. The methodology was a binary classification problem based on analysing conversational data between a participant with normal hearing and a participant with age-related hearing loss. These conversations took place in nine different conditions with varying background noise, and levels of hearing aid support for the participant with hearing loss. The ASR model used was Whisper by OpenAi™. The extracted features from the data included frequency characteristics, spectral attributes, power measures, pauses, floor-transfer offsets, Mel Frequency Cepstral Coefficients, and principal components derived from the first and final encoder block outputs of the Whisper model. The impact of the conditions and the features on the classification accuracies were analysed in conjunction to evaluate an optimal future method to detect hearing loss from speech data. The results suggest that it is possible to classify age related hearing loss with speech data, the top accuracy achieved was 78.4%. The background noise and the level of hearing aid support affected the outcome of the classification accuracy, as well as the selected features. The most influential features were related to power, frequency, and the Whisper encoder block outputs. This thesis does not incorporate control group conversations between participants with normal hearing, and the data set is limited to 24 participant pairs. To achieve more reliable and generalisable results, the inclusion of a control group and a larger data set is essential.}},
author = {{Bergström, Sanna and Ekre, Fremja}},
language = {{eng}},
note = {{Student Paper}},
title = {{Classification of Age-Related Hearing Loss using Conversational Speech Biomarkers}},
year = {{2025}},
}