Leveraging Modality-Specific Representations for Audio-Visual Speech Recognition via Reinforcement Learning.