Measurement scales, psychophysical models, and reliability statistics in voice quality perception

Abstract

Perceptual judgments of voice quality can be given on discrete or continuous scales. Discrete scales are typically described as consisting of “equal appearing intervals,” and continuous judgments may be given on visual analog scales (VAS) or via direct magnitude estimation. Interrater agreement on discrete scales is highly variable and often poor, though there is evidence that application of psychophysical principles (i.e., perceptual variability, response bias) can increase reliability. There is some evidence that interrater reliability may be higher with continuous scales, though the influences of psychophysical assumptions and measurement properties on the reliability of continuous scales has not been analyzed particularly thoroughly. The present work considers the effects of different assumptions about underlying perceptual and decisional processes on the measurement properties and reliability of perceptual voice quality scales. An adaptation of Stevens’ power law combined with different decision rules provide a model of the mapping between the acoustic properties known to influence voice quality (e.g., H1-H2, HNR) and discrete or continuous clinical judgment scales (e.g., “breathiness,” “roughness”). The effects of variation in model parameters on reliability and agreement statistics will be analyzed, and novel agreement statistics based on the properties of ratio-scale judgments will be explored.