Scaling analysis of consonant confusion matrices: An acoustic account

Abstract

Multidimensional scaling analysis of consonant confusion matrices has been frequently used as a means to derive the underlying perceptual representation of speech sounds. In many studies, dimensions yielded by such analyses have been interpreted in terms of binary or tertiary linguistic features. The results of these studies, however, have not converged to a single system of perceptual features. Soli and Arabic [J. Acoust. Soc. Am. 66, 46–59 (1979)], argued that an acoustic‐ rather than a feature‐based interpretation may provide a better account of multidimensional scaling solutions. The current study was undertaken to explore further the adequacy of such an interpretation by providing a quantitative description of the relations of acoustic properties of consonants to their respective projections on dimensions yielded by multidimensional scaling analysis. The material for acoustic and statistical analyses consisted of the speech stimuli and confusion matrices of Wang and Bilger [J. Acoust. Soc. Am. 54, 1248–1266 (1973)]. Stimuli included 24 CV and 19 VC syllables, the vowels being either /i/, /a/, or /u/. For each syllable, acoustic analysis yielded an ensemble of spectral, durational, and energy measures. Consonant confusion matrices were analyzed using the SINDSCL procedure. The results of regression analyses of the acoustic parameters and the perceptual dimensions obtained by the scaling analyses will be discussed in terms of their implications for an acoustic model of consonant confusions. [Work supported by NIH.]