Abstract
This article proposes a new technique for skimming multimedia content such as video mail, audio/visual data in blog sites, and other consumer-generated media. The proposed method, which is based on the automatic extraction of emphasized speech, locates emphasized portions of speech with high accuracy by using prosodic parameters such as pitch, power, and speaking rate. As the method does not employ any speech recognition technique, it enables a highly robust estimation in noisy environments. To extract emphasized portions of speech, the method introduces a metric, “degree of emphasis,” which indicates the degree of emphasis of each speech segment. Given an article, the method computes the degree of emphasis for each speech segment in it. When a user requests a skimming of the article's content, the method refers to the user-specified “skimming rate” to collect the emphasized segments. Preference experiments were performed in which participants were asked to select either the skimmed contents created by our method or those created using a fixed interval approach. The preference rate of our method was about 80%, which suggests that the proposed method can generate proper content skimming.
Acknowledgments
We thank Katsuhiko Ogawa, Director of NTT Cyber Solutions Laboratories, for encouraging us to write this article. We have had helpful discussions with Osamu Mizuno, Broadband Business Development Division, Corporate Business Headquarters, NTT EAST, Junji Takeuchi, Visual Communications Division, NTT Bizlink, and the office support staff of Megumi Machiguchi, Kaori Takei, and Harumi Matsuura.
Notes
1Cepstrum is a standard spectral parameter in the speech processing domain and is obtained by applying inverse-FFT to log area arithmic domain power spectrum (CitationBogert, Healy, & Tukey, 1963).