haskins logo
A-158 Variability and error in speech production
Louis Goldstein, NIH

Research Goals. This project addresses the relationship between 'normal' token-to-token variability in the production of phonetic units and tokens that can be characterized as containing ‘errors.’ Traditionally, speech errors are thought to occur when speakers produce something that was not intended through a categorical mis-selection of units during planning (e.g., 'fonal phonology' for 'tonal phonology'). Systematic studies of such sublexical errors have been largely carried out using phonetic transcription without any investigation of speech articulator activity. Conversely, studies of speech kinematics (and its variability) typically exclude tokens that are perceived as containing 'errors.' However, recent studies of articulatory kinematics during error-producing tasks have discovered a class of productions that sound like errors, but that do not involve categorical mis-selection of units. Rather, although an intended unit’s vocal tract constriction is correctly produced, it is accompanied—to various degrees across tokens—by an unintended constriction (an 'intrusion'). Thus, unintended constrictions represent a neglected source of token-to-token variation, and provide support for the hypothesis that errors and variation are more intimately linked than previously thought. Since the factors causing token-to-token variation are themselves poorly understood, Louis and colleagues propose to investigate patterns of variability and their relation to errors with three specific aims, for which they will:

     • manipulate experimental factors that should contribute to making an
utterance ‘hard to say,’ and measure their effect in three tasks: planning time (RT), articulatory variation in a speeded production task, and errors (as operationally defined) in a repetitive task, uncovering the patterns of correlations of these measures;
     • test a dynamical account of errors as shifts in the behavior of pairs of
coupled planning oscillators from more complex to simpler frequency ratios, with associated changes in fluctuations, i.e., variability; and
     • examine the acoustic and perceptual consequences of intrusions to test
the possibility that what have been called errors are simply instances of variation extensive enough to shift perceptual categories.

Relevance: Diagnoses of different speech disorders are often based on the clustering of error types and on the perceived variability of patients’ speech. The analyses will provide a sound theoretical and empirical basis for such acoustic and perceptual evaluations.

Current Status.  The proposal was funded for five years beginning 9/1/07. Total costs for the first year are $655,137.