A Data Augmentation Technique for Automatic Detection of Chewing Side and Swallowing

Poor quality of eating behavior is known to have adverse effects on health. With a view to promoting health, this study examines a monitoring system for eating behavior that uses a convenient microphone. We previously performed automatic detection of masticatory balance and swallowing using two-channel microphone recordings and the Hybrid CTC/Attention Model to detect the quality of eating behavior. In this paper, we propose an N-gram based data augmentation technique using a large amount of weakly labeled data to improve the accuracy of automatic detection. The application of this method to the Hybrid CTC/Attention Model resulted in improved detection performance. Moreover, the performance of open foods not included in the training data was shown to be similar to that of closed foods.