Prosodic parallelism as a cue to repetition disfluency

Repetition disfluencies are among the most frequent type of disfluency in conversational speech, accounting for over 20% of disfluencies, yet they do not generally lead to comprehension errors for human listeners. We propose that parallel prosodic features in the REP and ALT intervals of the repetition disfluency provide strong perceptual cues that signal the repetition to the listener. We report results from a transcription analysis of repetition disfluencies that classifies disfluent regions on the basis of prosodic factors, and preliminary evidence from F0 analysis to support our finding of prosodic parallelism. 1. Acoustic-prosodic correlates of disfluency Disfluency occurs in spontaneous speech at a rate of about one every 10-20 words, or 6% per word count [17], yet this interruption of fluent speech does not generally lead to comprehension errors for human listeners. Recent research has shown that important cues to disfluency can be found in the syntactic and semantic structures conveyed by the word sequence, and in the phonological and phonetic structures signaled by acoustic features local to the disfluency interval. These cues identify the components of the disfluent region--the reparandum (REP), edit phrase (EDIT), and alteration (ALT)--and their junctures. Work on automatic disfluency detection has shown that the most successful approach combines both lexical and acoustic features, with explicit models of the lexical-syntactic and prosodic features that pattern systematically with disfluent intervals [1,6]. Of the acoustic-prosodic correlates of disfluency, the postreparandum pause (filled or unfilled) has been studied the most extensively. Nakatani & Hirschberg’s [12] detailed acoustic and classification studies examine duration, F0 and energy, and also report unusual patterns of lengthening, coarticulation, and glottalization near the interruption point of a disfluency. In this paper we examine the nature of prosodic correlates of disfluency in the characteristic patterns of F0, duration and energy that identify and distinguish among various types of disfluency involving word repetition. There are distinct types of disfluency that can be characterized in terms of their form and function. Shriberg [16, 17] classifies the disfluencies of the Switchboard corpus into six categories: filled pause ("uh" and "um"), repetition (of one or more words, without correction), substitution (repetition of zero or more words, followed by the correction of the last word in the disfluent interval), insertion, deletion, and speech error. Other work identifies abandonment (fresh start) disfluencies, in addition [6,11,18]. These distinct types of disfluency may be caused by different psychological processes. Levelt [9] suggests that corrections of a single word may result from monitoring of the phonetic plan, while corrections that involve repair or abandonment of an entire phrase may result from monitoring of the pre-syntactic message. Clark and Fox Tree [3] and Clark & Wasow [4] propose a different psychological account for filled pause and repetition disfluencies. In these accounts filled pauses like “uh” and “um” are phonological words that are used by the speaker to signal a delay in the preparation of the upcoming speech. Repetition disfluencies occur when the speaker makes a premature commitment to the production of a constituent, perhaps as a strategy for holding the floor, and then hesitates while the appropriate phonetic plan is formed. The continuation of speech is marked by “backing up” and repeating one or more words that precede the hesitation, as a way of restoring fluent delivery. Henry and Pallaud [7] support the findings of Clark & Wasow [4] by demonstrating that morphological, syntactic, and structural features strongly differentiate repetition disfluencies from word fragment disfluencies. Clark & Wasow [4] note that repetition disfluencies are four times as common as repair disfluencies; they suggest that a small number of repetition disfluencies may be "covert repairs" [9], but that most repetitions are more closely related to filled pause disfluencies than to speech repairs. The acoustic-prosodic features that serve to cue disfluency vary according to the type of disfluency. Levelt & Cutler [10] observe a contrastive emphasis on the repair segment of an error-correcting disfluency, manifest in increased F0, duration and amplitude. Shriberg [15] and Plauché & Shriberg [13] find that F0 contours, word durations, and the distribution of pauses serve to differentiate among three types of repetition disfluencies. Shriberg [15] describes repetition disfluencies that signal covert repair as having a characteristic reset of the F0 contour to a high, phrase-initial value at onset of the alteration. Similarly, Savova and Bachenko [14] propose an “expanded reset rule,'' according to which “alteration onsets are dependent on both reparandum onsets and reparandum offsets,” echoing the observation of Shriberg [15] that when speakers modify the duration of a repeated word in a repetition disfluency, “they tend to do so in a way that preserves intonation patterns and local pitch range relationships.” In our study of prosody and disfluency in the Switchboard corpus of conversational telephone speech, we observe parallelism in the prosodic features of the REP and ALT phases as characteristic of some repetition disfluencies. Highly similar F0 patterns express a parallel intonation structure that cues the relationship between the REP and ALT for the majority of repetition disfluencies we have observed. We propose an extended typology of repetition disfluencies in this paper, based on prosodic comparison of REP and ALT. Section 2 describes the methods of our transcription study of disfluency in Switchboard, and section 3 presents frequency data on five types of repetition disfluency that are prosodically distinguished based on a comparison of the prosodic features of the REP and ALT intervals. Section 4 reports on preliminary quantitative evidence from F0 data to support our analysis based on perceptual transcription.