Source localisation with recurrent neural networks

We address the far-field source localisation problem. A new neuromimetic approach is presented. Instead of considering time as a second dimension of the input space, time is implicitly encoded in the structure of a recurrent neural network. This allows us to process temporal data such as the temporal propagation delays between the sensors of a linear array. Recurrent neural networks are encompassed in a two stage decision process. Locally, specialised neural networks are in charge of detecting acoustical sources in small angular sectors. Globally, an integration system locates accurately the direction of arrival of the signals.