An Analysis of the Variance of Diffusion-based Speech Enhancement