When a large number of documents, e.g. bank cheques, have to be authenticated in a limited time, the manual verification of, say the authors' signatures, is often unrealistic. This led to the development of a wide range of automatic off-line signature verification systems. However, the value of such a system is rarely demonstrated by conducting a subjective test. We recently developed a novel off-line signature verification system [2, 3] that uses features that are based on the calculation of the Radon transform (RT) of a signature image. Each writer's signature is subsequently represented by a hidden Markov model (HMM). This paper is an extension of [3] and illustrates the value of our system by showing that it outperforms a typical human being. We conduct an experiment on a data set that contains 765 test signatures (432 genuine signatures and 333 skilled forgeries) from 51 writers.