부산대학교 도서관

This paper presents e-Breath, a method, and system for monitoring breath using a wearable microphone connected to the smart phone. In addition, we investigate the performance of single feature models such as Mel Frequency Cepstral Coefficient and Grammatone Frequency Cepstral Coefficient for breath detection and propose a simple yet effective feature fusion model to improve the breath detection accuracies. Experiments on a dataset, which contains over 5,700 breath events and significant noises, collected from 16 persons worn the mobile devices several hours, have demonstrated that breath can be detected with the detection rate of 97% for individual evaluation, and over 92% for subject independent evaluation, which is improved from 5% to 8% compared to single feature models. e-Breath is highly potential for healthcare applications that acquire breath information for the diagnose and treatment of respiratory diseases.