Vowel normalization: Differences between vowels spoken by children, women, and men

Abstract

Using the data of Peterson and Barney [J. Acoust. Soc. Am. 24, 175–184, (1952)] we have sought transformations of the frequency scale (e.g., Mels) and rules of combination of values of the fundamental (F0) and the formants (F1, F2, and F3) (e.g., F2/F1, F3/F2) that will eliminate differences between talker groups (children, women, and men) while maintaining differences between vowel categories. The ratios of the between- to the within-vowel category variation as well as the minimum separations between vowel categories, scaled by the within-category variation, were calculated. Of the many frequency transformations and combination rules examined, the vector [log (F1/F̂), log (F2/F̂), log (F3/F̂)], where F̂ = (F1⋅F2⋅F3)1/3, provides the best clustering of the vowels and nearly eliminates differences between talker groups. This suggests that vowels are characterized by their spectral shapes when plotted on log axes. Thus the usual range of differences between children, women, and men in the absolute location of the spectrum along the log-frequency axis appears unimportant. [Supported by Grant NS 03856 from NINCDS.]