Classification of ingestion sounds using Hilbert-Huang transform

Abstract

Automatic classification of food ingestion gives a precise and objective solution for dietary monitoring which is an active research area. In this study, we aim to classify ingestion sounds of the six different food types recorded from the throat microphone. We observe that these records show a different energy distribution than normal speech signals. To reveal the characteristics of intake signals, we prefer a model that could reflect the energy distributions. Using the Hilbert-Huang transformation, we decompose the signal on the local time-scale. As a result of this hierarchical decomposition, zero-crossing rates and short-term energies are calculated for each component. These feature sets are then classified using the support vector machine classifier. After the experimental studies, a classification accuracy of 72% is obtained for the six-class classifier that indicates the proposed methodology is promising for further studies.