Automated Sound Analysis System for Home Telemonitoring using Shifted Delta Cepstral Features

With an ageing world population and a corresponding demand for aged care, interest in the development of home telemonitoring systems has increased greatly in recent years. Automated sound analysis systems have been considered as an alternative for video monitoring in the interests of privacy. This paper investigates the use of frequency domain features, namely the Mel frequency cepstral coefficients (MFCC), in identifying and monitoring sounds of daily activities of elderly persons. A Gaussian mixture model (GMM) is used as the back-end system classifier. We also include a new compact feature set, the shifted delta cepstrum (SDC), improving our results. This model achieves a classification accuracy of 91.58%, distinguishing between 19 different real-world sounds.