Improving Generalization of Deep Neural Network Acoustic Models with Length Perturbation and N-best Based Label Smoothing.