Results are reported of a pilot experiment in the use of hill climbing to derive the numerical parameters used in a rule-based approach to segmentation and labelling of continuous speech. Starting from two sorts of symbolically expressed rules, segment rules and feature rules, the user may nominate any numerical field for hill climbing, together with desired mean rate of change. Given a suitable corpus of positive and negative examples, the system then optimises the values of these fields with respect to overall labelling performance. An open test of context-specialised versions of an existing rule for identifying nasals showed a roughly 50% reduction in error rate.