Discriminative Manifold Learning For Automatic Speech Recognition