Analysing Diffusion-based Generative Approaches Versus Discriminative Approaches for Speech Restoration