An algorithm for establishing more representative initial reference data in a speaker recognition system is presented, together with five different methods for updating the reference data. These methods have been tested on a text-dependent speaker verification system and the results are presented.