We propose an unsupervised and universal method for speaker roles classification. We extract features from turns, speech segments and both of them for speaker role classification. We propose a role clustering algorithm based on maximizing inter-cluster distances.