Activity: Talks and presentations › Conference presentations
Spherical microphone arrays enable the directional decomposition of a sound field over all propagation angles and are therefore a valuable analysis tool in enclosures. At the same time, deep neural networks have shown promising performance in acoustic source localisation tasks, especially in challenging acoustic scenarios. Combining neural network source localisation with spherical array signals can prove extremely beneficial in existent applications. However, classical neural networks typically exploit one- or two-dimensional data correlations. This can cause the network estimation to be prone to errors in wave propagation direction and limiting the array to specific directions, since the network does not guarantee rotational invariance. In this study, we examine a spherical graph neural network architecture for direction of arrival estimation trained on both simulated and measured sound field features. The spherical graph neural network can capture spherical correlations and can warrant that they are held approximately invariant for rotations. The network's performance is preliminarily investigated under anechoic conditions with the respective measured spherical array impulse responses.