Semi-structured capture and display of telephone conversations

Speech dominates day-to-day communication, but despite technological advances, ordinary conversations have remained outside the realm of computer-supported work. This thesis addresses semi-structured audio, a framework for medium-specific information retrieval of stored voice that does not rely upon knowledge of the actual content. Instead, structure is derived from acoustical information inherent in the stored voice and augmented by user interaction. To demonstrate semi-structured audio, two software applications were constructed: the Listener, a tool for capturing structure while recording telephone conversations, and the Browser, for subsequent browsing of speech fragments. The Listener segments the audio signal, using changes in who is speaking to identify conversational turns, and pause detection to identify phrase boundaries. This conversational structure is dynamically determined during the phone call, and presented in a retrospective display that provides flexible capabilities for marking segments of interest. The Listener and Browser make use of the ChatViewer widget, which supports the display of, and user interaction with, collections of sound and text items called chats. This semi-structured approach makes it practical to retain and access large amounts of recorded speech. Thesis Supervisor: Chris Schmandt Title: Principal Research Scientist 'This work was supported by Sun Microsystems, Inc.