A Study of the Scale Invariant Signal to Distortion Ratio in Speech Separation with Noisy References