Direction of arrival estimation utilizing deep neural networks with dual channel microphones under noise environment

Abstract

Under noise condition, it is hard to estimate accurate direction of arrival(DOA). In this research, a deep neural network(DNN) is applied to improve accuracy and noise robustness. In addition, this work utilizes dual channel microphones with short interval and fan noises from drone are regarded as the noise source. This work includes random data generation, feature extraction, data augmentation, network selection and label smoothing for sound localization(DOA estimation). To solve data deficiency problem, data generation and augmentation for DOA estimation are suggested. Previous works applied DNN to DOA estimation utilized only signal processed data set to train network. This paper is including results of signal processed data set and recorded data set. To solve data deficiency issue in deep learning research, this work suggested data augmentation fit to DOA estimation problems, such as flipping, phase rotation, masking, noise addition. In addition, special label smoothings for localization are suggested to minimizing angular error. In conclusion, it shows 100% accuracy for clear data set, and achieved 61% accuracy for low SNR data set(-34 dB).