Sparse modeling of neural network posterior probabilities for exemplar-based speech recognition