Demo: Chewpin: a wearable acoustic device for chewing detection Yang Chen sonnechen95@gmail.com Division of Industrial Design, National University of Singapore Singapore Zhitong Cui zhitongcui@zju.edu.cn Zhejiang University Hangzhou, China Ching Chiuan Yen didyc@nus.edu.sg Division of Industrial Design and Keio-NUS CUTE Center, National University of Singapore Singapore ABSTRACT Diet intervention has emerged as a promising strategy in obesity prevention and treatment. Existing research predominantly focused on macronutrient intake and food quantity restriction in diet ma- nipulation. Eating habit is difcult to change due to its highly habit- ual nature; therefore, it is essential to automatically detect eating behavior and provide real-time intervention in unhealthy eating patterns. In this study, we explored the possibility of designing Chewpin, an easy-to-be-implemented and socially acceptable de- vice for capturing eating behavior(i.e., chewing and swallowing) in a controlled environment. We implemented a convolutional neural network (CNN) for data classifcation. Overall, our system achieved a promising accuracy of eating recognition of 98.23% on the test set. In the future, we will evaluate its usability and feasibility in real-life eating practices and use this system as a technical tool for problematic eating intervention. CCS CONCEPTS · Human-centered computing → Interaction devices. KEYWORDS eating detection, acoustic sensing, wearable device, CNN ACM Reference Format: Yang Chen, Zhitong Cui, and Ching Chiuan Yen. 2021. Demo: Chewpin: a wearable acoustic device for chewing detection. In Adjunct Proceedings of the 2021 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2021 ACM International Symposium on Wearable Computers (UbiComp-ISWC ’21 Adjunct), September 21–26, 2021, Virtual, USA. ACM, New York, NY, USA, 2 pages. https://doi.org/10.1145/ 3460418.3479290 1 INTRODUCTION Overweight and obesity have become one of the most signifcant public health issues in societies [4]. Accumulating evidence sug- gested that macronutrient intake is one of the leading causes of unhealthy weight gain [3]. Thus, researchers strived to monitor eat- ing activity as a means to gather information on problematic eating and implement dietary intervention to help people regularize eating Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). UbiComp-ISWC ’21 Adjunct, September 21–26, 2021, Virtual, USA © 2021 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-8461-2/21/09. https://doi.org/10.1145/3460418.3479290 habits and take back control of their health. One major challenge for eating intervention is to understand when people eat. Researchers in the area of eating event detection have proposed several tech- niques to determine eating-related behaviors. For instance, audio sensors were used for collecting acoustic information of chewing and swallowing in the ear canal or on the throat [1, 2, 6], camera for frst-person image analysis of eating behavior [5, 7, 9], wrist-based eating gesture recognition [8, 10]. Although these systems have been widely explored in eating detection, they have several practi- cal limitations such as obtrusive, privacy-invasive, and not social acceptability in real-life eating practice. In this preliminary study, by weighing the pros and cons, we explored the usability of acoustic sensing, an easy to be implemented and socially-acceptable tech- nique for eating detection. Two major design considerations should be met: 1. It should be suitable to be worn in the real world: comfort- able and socially acceptable. 2. It should capture information-rich features from audio and accurately distinguish eating episodes in a given scenario. With these concerns in mind, we designed a wear- able acoustic device, Chewpin, as a technical tool for audio data collection. Then, we implemented this device in the eating sound collection in a lab-controlled environment and extracted features to train a convolutional neural network (CNN) to classify eating and non-eating behavior. To the best of our knowledge, no study has applied CNN for acoustic eating sound detection before. The accuracy of our system showed a promising result of 98.23% on the test set. The contribution of this ongoing exploratory study are: • Design and implement a wearable acoustic sensor-based device that can capture eating behavior (i.e., chewing, swal- lowing). • Develop and evaluate an eating detection classifer model using CNN based on a set of acoustic features. 2 SYSTEM DESIGN 2.1 Hardware device Acoustic microphone A dual-microphone expansion board (ReS- peaker 2-Mics Pi HAT) was applied to capture eating sound. The board is developed based on WM8960, low-power stereo codes, and two microphones which are capable of collecting nuanced acoustic sounds supporting a max 48kHz sampling. This device has been used in various AI assistant and voice interaction applications. Mini computer Raspberry Pi Zero W was chosen as a tiny, low-cost mini-computer with on-board wireless LAN and Bluetooth 4.1 for hardware prototyping. This mini-computer is compatible with a dual-microphone and is feasible to be implemented in real-living conditions as a wearable device for eating data collection (Figure 1). 11