Improving the Decoding Efficiency of Deep Neural Network Acoustic Models by Cluster-Based Senone Selection.