To carry on effective, natural, spoken dialogue,computers will need to do more than recognize and interpret spoken words -- they will also need to be sensitive to the prosodic information encoded in how the words are spoken. To investigate the hypothesis that detecting and exploiting prosodic cues can help computers guide spoken dialogue, this case study focuses on an educational task that combines intrinsic national importance with compelling methodological advantages: listening to a child read aloud, and providing spoken assistance. The research focusses on improving four aspects of dialogue -- taking turns, handling speech repairs, preventing dialogue breakdown, and modelling the speaker. In the reading task, these aspects include detecting a number of pedagogically significant events, such as when readers complete a passage, correct themselves, or encounter difficulty in identifying a word or comprehending a passage. The goal is to use prosodic cues to help detect these events in order to make the dialogue between student and computer more effective in achieving its educational objectives . Expected outcomes include improvements in the automated reading assistant, the discovery of robust prosodic phenomena, methods for detecting them, and principles for using them to achieve graceful, effective spoken dialogue between humans and computers.
