Gated Recurrent Fusion With Joint Training Framework for Robust End-to-End Speech Recognition