This paper deals with extremely similar-sounding, unrelated speakers ('voice twins') and presents an automatic approach to voice twin discovery applied to different speaker databases. An automatic speaker recognition system relying on perceptually relevant phonetic features including formants and a tuned clustering algorithm DBSCAN was used to group recordings within diverse datasets. 18 voice twin pairs selected from 2-speaker clusters were evaluated by 50 listeners in a 2-alternative forced choice experiment. Same/different decisions and confidence ratings were collected for same-speaker, random different-speaker and voice twin comparisons. Listeners were unable to differentiate between the candidate voice twin pairs much better than chance level while they performed well (80% accuracy) for random same- or different-speaker comparisons indicating the voice twin speakers were perceptually very similar. The implications and forensic relevance of identifying voice twins are discussed.