A Survey of Deep Neural Network in Acoustic Direction Finding

Abstract

Direction of Arrival (DoA) estimation has importance in many industries such as speech enhancement, spatial audio coding, radio frequency and radio telescope. Deep Neural Network (DNN) has find its way into DoA applications along with the well-known methods such as subspace-based or time difference of arrival methods, which opens-up the data-driven approach towards estimating the DoA. This paper first surveys different DNN architectures and their supporting methods and datasets that are used for estimating DoA in different scenarios. Then a promising architecture based on convolutional recurrent neural network (CRNN) is re-presented on the Spatially Oriented Format for Acoustics (SOFA) dataset, where the average error rate of 9.68° has been achieved.