Shaping Spoken Input in User-Initiative Systems

A spoken dialog system performs best when users speak within the grammar that the system understands. We conducted a simple study to investigate how easily users can be persuaded to speak to a system using a restricted, less-thannatural- language input style. In a Wizard-of-Oz setting, users of a spoken dialogue system for information access were given brief instructions to “speak simply” to the system. During their interactions, conversational or complex input was rejected by the system while simpler, “just-the-facts” input was accepted. We found that all users were able to adapt their language to successfully complete tasks, and participants’ post-experiment comments showed that they were consistently mindful of the form of acceptable input. These results will support further investigation into more precise shaping of user input, leading to more effective and efficient human-machine speech interaction.