Self-organizing maps (SOM) and multilayer perceptron (MLP) neural network approaches are applied to the evaluation of spatial discrimination of real and virtual sound sources. Neural networks are trained with localization cues computed using a binaural model. The ability of the models to simulate human perception of spatial sound is analyzed. Both SOM and MLP showed relatively good ability for generalization when tested with test data from real sources not included in the training data set. Both models were also capable of describing localization blur of virtual sources though still further analysis is needed to find out how well our models correspond to real human auditory localization. The motivation of this study has been to search for methods and techniques to evaluate the quality of spatial sound reproduction