Audio-Based Onset Detection applied to Chewing Cycle Segmentation

In this paper we compare three onset detection algorithms for acoustic chewing cycle detection, which is a basic step in eating detection and automated dietary monitoring. We introduce a spectral flux algorithm that uses the spectrogram of a chewing sequence to compute a novelty function. Furthermore, beat tracking, in particular the notion of a predominant local pulse is introduced. We compare the two algorithms to a baseline energy-based segmentation in a chewing dataset with seven participants consuming pieces of six different foods, including in total 9818 annotated chewing cycles. Best performance was achieved for the beat tracking algorithm with 83% F-measure after leave-one-participant-out cross validation.