Acoustic event characterization for service robot using convolutional networks

Abstract

This paper presents and discusses the creation of a sound event classification model using deep learning. In the design of service robots, it is necessary to include routines that improve the response of both the robot and the human being throughout the interaction. These types of tasks are critical when the robot is taking care of children, the elderly, or people in vulnerable situations. Certain dangerous situations are difficult to identify and assess by an autonomous system, and yet, the life of the users may depend on these robots. Acoustic signals correspond to events that can be detected at a great distance, are usually present in risky situations, and can be continuously sensed without incurring privacy risks. For the creation of the model, a customized database is structured with seven categories that allow to categorize a problem, and eventually allow the robot to provide the necessary help. These audio signals are processed to produce graphical representations consistent with human acoustic identification. These images are then used to train three convolutional models identified as high-performing in this type of problem. The three models are evaluated with specific metrics to identify the best-performing model. Finally, the results of this evaluation are discussed and analyzed.