Detection of supralexical dysfluencies in a text read by children who stutter

Abstract

When asked to read a text, people who stutter produce dysfluencies that can be divided into two types. These are (a) those dysfluencies that mainly influence production of individual lexical items and (b) those involving, either alone or in combination, omitting words, inserting incorrect words and repeating phrases. In case (a), the speech breakdowns are termed lexical dysfluencies (LD). LD include word and part-word repetitions, prolongations, and broken words. The dysfluencies in (b) are termed supralexical (SD). This class comprises interjections, revisions, incomplete phrases, and phrase repetitions. If SD and LD are not distinguished, then the way certain dysfluent words should be categorized is inherently ambiguous. For instance, there is no a priori way of deciding how to categorize an LD that occurs within a group of words comprising an SD. The proposed solution to this problem involves locating and processing SD before LD. Doing this allows any LD that occurs within a group of words that can also be designated as an SD to be assessed and removed from further consideration prior to location of isolated LD (the LD that occur within an SD are, then, subordinate to the SD). A computer-based parser that locates SD in transcriptions of read text is described. Its performance is compared with that of human judges.