3D Direction-of-Arrival estimation through spherical acoustic intensity and spherical structured deep learning network

抄録

Neural networks have proven valuable for estimating the Direction-of-Arrival (DoA) of acoustic signals since they are capable of overcoming the accuracy and robustness inherent to conventional estimation methods when dealing with acoustic phenomena. This paper presents a system based on the processing of the acoustic intensity formulated in spherical coordinates. Due to the omnidirectional and spherical structure of this type of features, a spherical convolutional neural network architecture is used to estimate the DoA by means of a regression task. A series of experimental tests based on angular error have been performed to examine the accuracy as a function of reverberation and noise, demonstrating the degree to which the proposed method offers competent robustness compared to some state-of-the-art methods.