The process of speaker recognition is generally based on modeling the characteristics of each speaker. An interesting method for modeling consists in representing a new speaker, not in an absolute manner, but relatively to a set of well trained speaker models which constitutes the new representation space. This paper addresses the task of finding a good representation space for speaker identification. It describes a representation space built either by clustering speakers or by selecting an optimal subset of them. In this representation space, speaker location is then performed by the anchor models technique. We present experimental results and compare them with GMM-based results. We show that clustering and subset selection give good representation spaces. With a little amount of training data, identification by location in a space of virtual voices performs much better than GMM.