Convolutional neural network-based direction-of-arrival estimation using stereo microphones for drone

Abstract

Recently, with the development of drone technology, various researches applying drones have been conducted. Among them, sound source localization for the drone is an important research topic because it can be utilized to find the person in an emergency. However, microphones mounted on the drone capture the ego-noise such as wind and fan noise generated from propellers, resulting in the extremely low signal-to-ratio condition. In this paper, we propose a method of direction-of-arrival (DOA) estimation using stereo microphones for the drone. To cope with the ego-noise, parametric multi-channel Wiener filter is used. After that, power level-based features are fed into the convolutional neural network model to classify the DOA of the desired speech signal. To evaluate our method, we used mean square error between the estimate and reference DOA as a metric. In our experiments, we recorded all utterances used to train and test in a real environment, and the result showed that the proposed system can be used to localize the sound source for the drone.