Speech Modeling and Enhancement Using Diffusion Maps