Conventional speech recognition approaches usually wait until the user has finished talking before returning a recognition hypothesis. This results in spoken dialogue systems that are unable to react while the user is still speaking. Incremental Speech Recognition (ISR), where partial phrase results are returned during user speech, has been used to create more reactive systems. However, ISR output is unstable and so prone to revision as more speech is decoded. This paper tackles the problem of stability in ISR. We first present a method that increases the stability and accuracy of ISR output, without adding delay. Given that some revisions are unavoidable, we next present a pair of methods for predicting the stability and accuracy of ISR results. Taken together, we believe these approaches give ISR more utility for real spoken dialogue systems.