Vowel spectra, vowel spaces, and vowel identification.

Abstract

Twelve Dutch vowels, each pronounced by 50 male speakers, were analyzed in 18 filter bands comparable in bandwidth with the ear's critical band. By considering the sound levels (in decibels) in these filter bands as dimensions, with a principal-component analysis the 18 dimensions per sound were reduced to four factors which together explain 75% of the total variance. The configuration of the average vowels in the factor space appeared to be highly correlated with their configuration in the F1−F2 formant plane. After matching to maximal congruence, correlation coefficients along corresponding axes were 0.997 and 0.979. Machine vowel identification, based upon the position of the individual vowels in the four-dimensional factor space, resulted (after three pairs of related vowels were grouped together) in 98% correct identifications if correction was applied for personal timbre of the speakers' voices. Ten listeners, to whom the 600 vowels were presented as 100-msec segments, gave 86% correct responses in identifying the intended vowels. The confusions between the vowel types were basis for a multidimensional scaling (Kruskal) to construct a perceptual configuration of the vowels. In four dimensions the solution showed 2.3% stress. Perceptual configuration and factor configuration, maximally matched, had correlation coefficients along corresponding axes of 0.997, 0.995, 0.907, and 0.794, respectively.