The Trouble with "Articulatory" Pauses

Abstract

The historical provenance of a minimum cutoff point (of about 0.25 sec) for pauses in temporal analyses of speech production is associated with Goldman-Eisler's usage. Her rationale was the predominance of articulatory pauses at lengths shorter than 0.25 sec. Both phonotactic facts and empirical analysis of several corpora of readings disconfirm this predominance with respect to pauses 0.13-0.25 sec in length. The vast majority of these pauses are found to be psychological; they are determined by syntax, punctuation, rhetorical and expressive emphasis, poetic format, and stylistic pecularities. * The research on political speeches reported in the following has been supported by a Deutsche Forschungsgemeinschaft Habilitationsstipendium granted to Sabine Kowal.

Fluency Profiling System: An automated system for analyzing the temporal properties of speech

“…In addition to ameliorating time costs, an automated analysis provides substantial benefits in standardization, objectivity, and reliability. Previous work from different laboratories has been characterized by a high level of idiosyncrasy (e.g., in the short-pause identification threshold; Goldman-Eisler, 1968;Hieke et al, 1983;Jaffe & Feldstein, 1970;Kirsner et al, 2005). Likewise, the subjective nature of manual segmentation places an upper limit on the reliability even for the best trained operators.…”

“…Manual segmentation-that is, the separation of prerecorded speech segments into speech and pause components-has previously relied on arbitrary minimum pause durations that exclude potentially meaningful information (Goldman-Eisler, 1968;Hieke, Kowal, & O'Connell, 1983) while also being highly variable and demonstrating only low reliability across studies, across analysts, and across sessions (Oehmen, Kirsner, & Fay, 2010). The purpose of this article is to introduce a fully automated process for the segmentation and analysis of temporal speech and pause characteristics.…”

“…1; Demol, Verhelst, & Verhoeve, 2007;Kirsner et al, 2002). The two components of the bimodal pause distribution correspond to short pauses important for articulation (approximately less than 250 ms; Hieke et al, 1983) and longer pauses related to breathing, planning, retrieval, discourse coordination, and a variety of other cognitive functions (Kirsner et al, 2005). Both short and long pause duration distributions provide valuable information about fluency across different clinical populations and different social contexts.…”

“…Short articulatory pauses are not related to cognitive processes and should be treated separately; however, determining the criterion between articulation and cognition is difficult (see, e.g., Hieke et al, 1983). If we assume that the log-normal pause distribution is bimodal, how do we appropriately classify the pause segments as either short or long in order to minimize the classification error rate in a principled manner?…”

“…Traditionally, the pause durations are segmented from speech using a visualization program (e.g., Praat; Boersma, 2001) and by manually entering boundaries at speech-pause transitions. The threshold between the short-and long-pause duration distributions is then either estimated by hand or derived by using an ad hoc criterion (Goldman-Eisler, 1968;Hieke et al, 1983;Jaffe & Feldstein, 1970;Kirsner et al, 2005); however, the latter method typically ignores the distinction between short and long pauses (e.g., thresholds range from 100 to 1,000 ms), and the former can be error prone and unreliable. Additionally, both approaches fail to consider individual differences between speakers, such that while the 250-ms threshold employed by Goldman-Eisler appears to be a reasonable approximation on average, it results in substantial errors of classification for most individual speakers.…”