A Context-Aware Baby Monitor for the Automatic Selective Archiving of the Language of Infants

Understanding the language of a newborn is not an easy task, and parents use baby monitors to monitor the language of their baby and infer the baby's needs. Available baby monitors trying to identify the language of babies consider the information after the baby is crying, and according to pediatricians the sounds and movements before crying define what the baby wants. In this paper, we explore the automatic selective archiving of the movements and sounds infants make before crying to give parents information they could later use to reflect on these behaviors and determine what the infant wants. Following a user-centered design methodology we developed a capture and access tool that uses accelerometers to detect when the infant is moving, and an algorithm to detect when the infant is crying. The movement detection algorithm uses a motion shimmer sensor and computes the magnitude exerted over the three axes, and the cry detection algorithm uses a Multi-Band Spectral Entropy Signature (MBSES) and a Support Vector Machine (SVM) to detect sustained crying. To show the feasibility of the performance of our system under realistic conditions, we tested how our sound algorithm performs under noise scenarios. The results show our algorithm is accurate, 98% precision, and performs better than algoritms using the MFCC feature. We close discussing directions for future work.