Improving sequence-to-sequence speech recognition training with on-the-fly data augmentation.