Gaussian duration modelling in semi-continuous hidden semi-Markov models is used to improve the performance of an automatic speaker verification system tested on telephone quality speech. A performance improvement of 2.3% in equal error rate for a 12 digit string is reported. Cepstra, delta cepstra, energy and energy plus delta energy parameter sets are all shown to contain speaker discriminating information, and a combination of these parameters, using equal weightings improved the system performance from 2.3% to 1.3% for a 12 digit string.