The Feature Extraction, Modeling And Recognition Of Dietary Behavior Signal Based On Bone-conducted Microphone

With the living standards improving, people are giving more and more attentionto their own health. However, people can not always understand their health becauseof the constraints of time and place. As a result, m-Health or mobile health arised.This emerging concept has been widely popular because it can get rid of the time andgeographical contraints to monitor the human condition. It is one of the m-Healthmonitoring methods that the analysis and processing of dietary behavior signal basedon bone-conducted microphone considered starting from the eating behavior.In recent years, speech recognition has been developed rapidly, but also achieveda high recognition accuracy. But the dietary behavior signal based on bone-conductedmicrophone has a big difference with traditional speech signal both in time domainand frequency domain. Therefore, the feature extraction, modeling and recognition ofspeech recognition can not apply to the dietary behavior signal based onbone-conducted microphone. Given this situation, this paper firstly makes asystematic statistical analysis of dietary behavior signal based on bone-conductedmicrophone including the characteristics analysis of time and frequency domain, anddraw some peculiar conclusions of dietary behavior signal. On this basis, a moredetailed analysis is made for the relationship between the distinguishing informationof dietary behavior signal based on bone-conducted microphone and frequencycomponents. Then we propose a new feature extraction method based on state meanF-ratio and the extracted features SMFFCC can characterize bone-conducted dietarybehavior signal. In the modeling and recognition, we also give full consideration ofthe characteristics of bone-conducted dietary behavior signal and determined uniquemodel parameters for training and recognition through the results of statisticalanalysis and a large number of experiments.Although the SMFFCC features are proposed for bone-conducted dietarybehavior signal, they are not limited to this. The process of analysis and derivation ofSMFFCC is not restricted to both bone-conducted signal and dietary behavior signal,but the new features obtained by a universal approach may reflect the characteristicsof bone-conducted dietary behavior signal finally. Similarly, we can use this methodto analyze the distinction of some kinds of sound in the social life or natural worldand eventually get the SMFFCC features that can characterize the sound better.The corpus for bone-conducted dietary behavior signal covers23people about 500minutes of data. We conducted experiments on this corpus and obtained betterresults, which basically reached the requirements of practical applications. In addition,comparative experiments show that the error rate using the proposed feature based onstate mean F-ratio is reduced and also show the effectiveness of the new features.