Mockingjay: Unsupervised Speech Representation Learning with Deep Bidirectional Transformer Encoders