Multilayer neural networks for speaker normalization