Semi-Supervised Real-World Adaptation for Swallowing Sound Detection

Abstract

Regular screening of swallowing is essential for preventing choking and aspiration pneumonia, particularly in older adults. Although skin-contact microphones enable noninvasive monitoring, studies primarily rely on controlled laboratory recordings for training and evaluation, without sufficiently considering the characteristics of natural dining settings. To address this limitation, we propose a semi-supervised learning strategy that adapts a swallowing sound detection model to natural dining settings by leveraging labeled and unlabeled real-world recordings. Experiments on real-world recordings demonstrate that the proposed method outperforms supervised baselines, achieving an F1 score of 0.933. Results confirm the effectiveness of semi-supervised real-world adaptation for the reliable detection of swallowing sounds in natural dining settings.