The effect of utterance length on the estimation of the likelihood of a speaker has previously seen a brief treatment in past works. In many speaker recognition evaluations, the utterances were typically configured to have a relatively consistent length. This paper investigates the effect of varying enrollment and test utterance lengths on the score distributions and consequently their effect on performance. In addition, this study examines the mixing of a number of variable training length utterance trials in a single evaluation. To address this problem, a parametric solution using splines is proposed and is shown to significantly reduce the error in the conducted experiments.