Is speech enhancement pre-processing still relevant when using deep neural networks for acoustic modeling?