We summarise experiments carried out using a system-initiative spoken CALL system, in which permitted responses to prompts are defined using a minimal formalism based on templates and regular expressions, and describe a simple structural learning algorithm that uses annotated data to update response definitions. Using 1 927 utterances of training data, we obtained a relative improvement of 20% in the system's ability to react differentially to correct and incorrect input, measured on a previously unseen test set. The results are significant at p < 0:005.