Stability and Accuracy in Incremental Speech Recognition