Automatic detection of hearing loss from children's speech using wav2vec 2.0 features

This study explores the feasibility of employing machine learning to classify acoustic features of speech for detecting hearing loss in preschool children. Acknowledging the critical developmental impacts of early hearing loss identification and the challenges associated with traditional testing methods for this age group, we propose a novel, scalable approach leveraging automatic speech analysis. Using speech recordings from children with and without hearing loss, we used wav2vec 2.0 and ComParE feature sets to capture speech characteristics and compared LSTM, DNN, and XGBoost classifiers. Our findings reveal that these models can accurately distinguish between the speech of children with hearing loss and those with normal hearing, achieving up to 96.4% accuracy. This proof-of-concept study indicates the potential of using speech for early hearing loss detection, and a path toward non-intrusive, scalable screening tools that could significantly benefit early developmental outcomes.