|Abstract:||Exemplar-based approaches to phonology presuppose that each exemplar can be assigned to a particular point in some fixed-dimensional mathematical space, such that it has a measurable distance from all other exemplars. But the speech signal is inherently of variable length. A new technique from the machine learning literature, the string kernel, permits variable-length signals to be mapped to fixed-dimensional space, by computing the extent to which they contain the same (contiguous and non-contiguous) sub-sequences. In this article, the string kernel technique is applied to a computational implementation of a phonological exemplar-based learning system (PEBLS). An experiment is presented, using this model to test an exemplar-based account of phonologization.