We attempt to explain a decrease in recognition word error rate observed when users interacted over time with a spoken language system. We found no change in the language used (as measured by sentence perplexity), and only a small decrease in the number of out-of-vocabulary words. However, a behavior adversely affecting recognition, hyperarticulation, decreased over time. In addition, the acoustic match of hyper-articulated utterances to the system models also improved over time. We conclude that improvement in recognition was due to changes in speech rather than in language.