This paper addresses the task of text independent speaker recognition and in particular looks at capturing time sequence information within the modelling process itself. A recent extension to the popular Gaussian mixture model (GMM) is the segmental mixture model (SMM), and its advantages are thought to be more pronounced as more and more training data becomes available. Here this idea is examined along with a hypothesis on model size, model complexity and their dependencies on the quantity of available training data. Experimental results on a 2000 speaker database show that an SMM does offer better recognition results than a GMM once a threshold in the amount of training data has been reached.