Adversarial Speaker Disentanglement Using Unannotated External Data for Self-supervised Representation-based Voice Conversion.