GAN Based Audio Noise Suppression for Victim Detection at Disaster Sites With UAV

Abstract

This paper presents a noise suppression system designed for unmanned aerial vehicles (UAVs). Searching for people using robots is expected to become a useful tool for saving lives during disasters. In particular, because UAVs can collect information from the air, there has been much research in rescue support using UAVs equipped with cameras. However, a limitation of cameras is their difficulty in detecting victims who are concealed. To solve this problem, we propose the use of a listening device on UAVs to detect sounds created by humans. This device uses an on-UAV microphone to capture human voices, which often get mixed with the sound of the UAV's propellers. This mixing presents a major challenge in identifying human voices. In this paper, we introduce a method to suppress the UAV propeller sound noise from the mix, enhancing the clarity of the human voice. Suppression of UAV sound noise is performed by generating pseudo-UAV sound based on generative adversarial networks (GAN) and reducing the generated pseudo-UAV sound from the sound mixture. By conducting various types of experiments, we confirmed the effectiveness of our proposal. As a result, we established the feasibility of using UAV-based voice processing for victim detection at disaster sites.