An INDSCAL Analysis of the Miller‐Nicely Consonant Confusion Data

Data from the Miller‐Nicely study [J. Acoust. Soc. Amer. 27, 338–352 (1955)] of perceptual confusions among acoustically degraded consonants were reanalyzed by Carroll and Chang's INDSCAL procedure [Psychometrika (1970) (to be published)]. While the input to ordinary multidimensional scaling is a single matrix of interstimulus similarities, the input to INDSCAL consists of several such matrices, one for each subject or other data source. The output shows the coordinates of the stimuli on the dimensions of a “group stimulus space” and also indicates the importance, or weight, of each dimension to every source. The similarities for a particular source are simply related to interpoint distances in a “private space” obtained by applying that source's weights to the respective dimensions of the “group stimulus space”. In this application the matrices arising from confusions among 16 consonants under each of 17 degradation conditions (the data sources) served as input. Six dimensions, interpreted as “Voicing,” “Nasality,” “Voiceless Stops versus Voiceless Fricatives,” “Second Formant Transition,” “Sibilance,” and “Sibilant Discrimination” arose from the analysis. The relative importance of these dimensions was systematically related to the kind and degree of degradation the consonants had undergone. Implications of the results for perception of consonants are discussed.