Investigating the contribution of speaker attributes to speaker separability using disentangled speaker representations