Multi-Encoder Learning and Stream Fusion for Transformer-Based End-to-End Automatic Speech Recognition.