Deriving a strategy for synthesizing lengthening disfluencies based on spontaneous conversational speech data

Our overarching research project explores the usability of disfluencies in incremental spoken dialogue systems. This endeavor requires basic phonetic research on disfluencies in spontaneous speech corpora as to define strategies for synthesizing disfluencies in a meaningful way. In this paper, our current research focus lies in an investigation of disfluency-related lengthening as a promising time-buying strategy in synthesized dialogue [1][2]. We base our analyses on the results of a search tool aiming to automatically detect lengthening in spontaneous speech corpora occurring without adjacency to phrase boundaries or other disfluencies, i.e. standalone lengthening phenomena. We analyzed disfluency-related lengthening in the ”monomodal” half of the GECO corpus [3], with regard to their context, word class, syllable position and phone type. We then postulate a disfluency insertion strategy for synthetic speech that prioritizes lengthening phenomena based on the results obtained in our study.