Toward a Better Understanding of Deep Neural Network Based Acoustic Modelling: An Empirical Investigation