Contrastive Learning from Synthetic Audio Doppelgangers