Sequence-to-sequence Automatic Speech Recognition with Word Embedding Regularization and Fused Decoding.