Time-Compression of Speech in Information Talks Using Spectral Entropy

Speech is one of the richest sources of information in human communication. In several situations, for example, classroom lectures, speech data alone conveys the majority of pertinent information. Large speech archives of lecture material are increasingly available on the Internet, but the lack of effective audio skimming tools complicates the interaction between users and audio material. In this paper, we describe a novel method for time-compression of speech based on the information theoretic measure of entropy. We demonstrate the applicability of spectral entropy in identifying important segments in speech, enabling time-compression of speech for skimming. We present results from a user study which confirms the viability of our approach.