DoA Estimation of Room Reflections Using NN-Based MUSIC Algorithm

The estimation of direction of arrival (DoA) for coherent sound sources, such as direct sound and early reflections in a room environment, is crucial. However, most traditional algorithms, like the multiple signal classification (MUSIC) algorithm, require observed signals to be non-coherent and can only be applied to narrowband signals. To tackle this issue, this paper introduces a neural network (NN)-based broadband MUSIC algorithm. By substituting the covariance matrix in the MUSIC algorithm with a pseudo-covariance matrix estimated through the network, this method can overcome the limitation of locating coherent broadband sources. Simulation results have shown that this proposed approach can accurately estimate the DoAs of direct sound and first-order reflections in both noisy and reverberant environments.