Learning to Speak to a Spoken Language System: Vocabulary Convergence in Novice Users

A key challenge for users and designers of spoken language systems is determining the form of the commands that the system can recognize. Using more than 60 hours of interactions, we quantitatively analyze the acquisition of system vocabulary by novice users. We contrast the longitudinal performance of long-term novice users with both expert system developers and guest users. We find that novice users successfully learn the form of system requests, achieving a significant decrease in ill-formed utterances. However, the working vocabulary on which novice users converge is significantly smaller than that of expert users, and their rate of speech recognition errors remains higher. Finally, we observe that only 50% of each user’s small vocabulary is shared with any other, indicating the importance of the flexibility of a conversational interface that allows users to converge to their own preferred