Robust offline trained neural network for TDOA based sound source localization

Passive sound source localization (SSL) using time-difference-of-arrival (TDOA) measurements is a non-linear inversion problem. In this paper, a data-driven approach to SSL using TDOA measurements is considered. A neural network (NN) is viewed as an architecture constrained non-linear function, with its parameters learnt from the training data. We consider a three layer neural network with TDOA measurements between pairs of microphones as input features and source location in the Cartesian coordinate system as output. Experimentally, we show that, NN trained even on noise-less TDOA measurements can achieve good performance for noisy TDOA inputs also. These performances are better than the traditional spherical interpolation (SI) method. We show that the NN trained offline using simulated TDOA measurements, performs better than the SI method, on real-life speech signals in a simulated enclosure.