4th International Conference on Spoken Language Processing
Philadelphia, PA, USA
Conversational speech is notably different from read speech in several ways, particularly in the presence of disfluencies but also in the frequent use of a small set of words that mark the flow of the discourse. Disfluencies are sometimes viewed as a \problem" in language modeling, where most previous work has focused on written text. In this paper, we take the view that disfluencies provide information themselves. In particular, we give evidence that filled pauses serve different functions, including marking linguistic unit and restart boundaries, and signaling hesitation where the speaker wants to hold the floor. The different functions can be connected to similar functions of other words common in spontaneous but not written speech, and the particular function affects the word conditioning choices in a variable n-gram model. Thus, at least some of the idiosyncrasies of spontaneous speech can be viewed as a source of information for language modeling rather than an interruption in the linguistic structure.
Bibliographic reference. Siu, Man-hung / Ostendorf, Mari (1996): "Modeling disfluencies in conversational speech", In ICSLP-1996, 386-389.