Deep Learning of Articulatory-Based Representations and Applications for Improving Dysarthric Speech Recognition