ISCA Archive

Disfluency in Spontaneous Speech

Aix-en-Provence, France
September 10-12, 2005


Important and New Features with Analysis for Disfluency Interruption Point (IP) Detection in Spontaneous Mandarin Speech

Che-Kuang Lin (1), Shu-Chuan Tseng (2), Lin-Shan Lee (1)

(1) Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan
(2) Institute of Linguistics, Academia Sinica, Taipei, Taiwan

This paper presents a whole set of new features, some duration-related and some pitch-related, to be used in disfluency interruption point (IP) detection for spontaneous Mandarin speech, considering the special linguistic characteristics of Mandarin Chinese. Decision tree is incorporated into the maximum entropy model to perform the IP detection. By examining performance degradation when each specific feature was missing from the whole set, the most important features for IP detection for each disfluency type were analyzed in detail. The experiments were conducted on the Mandarin Conversational Dialogue Corpus (MCDC) developed by the Institute of Linguistics of Academia Sinica in Taiwan.

Full Paper

Bibliographic reference.  Lin, Che-Kuang / Tseng, Shu-Chuan / Lee, Lin-Shan (2005): "Important and new features with analysis for disfluency interruption point (IP) detection in spontaneous Mandarin speech", In DiSS-2005, 117-121.