The problem of identifying approximately duplicate records between databases is known, among others, as duplicate detection or record linkage. To this end, typically either rules or a weighted aggregation of distances between the individual attributes of potential duplicates is used. However, choosing the appropriate rules, distance functions, weights, and thresholds requires deep understanding of the application domain or a good representative training set for supervised learning approaches. In this paper we present an unsupervised, domain independent approach that starts with a broad alignment of potential duplicates, and analyses the distribution of observed distances among potential duplicates and among non-duplicates to iteratively refine the initial alignment. Evaluations show that this approach supersedes other unsupervised approaches and reaches almost the same accuracy as even fully supervised, domain dependent approaches.