State-of-the-Art Speech Recognition Using Multi-Stream Self-Attention with Dilated 1D Convolutions.