Vowel normalization using stop consonant loci.

Abstract

Vowel normalization techniques used in sociophonetics aim to eliminate the influence of vocal tract length differences on resonant frequencies so that speaker‐specific vowel configurations can be compared more precisely. In vowel‐extrinsic methods, a speaker’s F1‐F2 range is scaled up or down, and vowel spaces are aligned using the geometrical F1‐F2 mean. However, a speaker’s F1‐F2 mean may be systematically biased if his or her vowel configuration is affected by a set of parallel vowel shifts, e.g., fronting or raising of several vowels in a chain shift. To address this problem, an alternative method of deriving speaker‐specific scaling factors was tested based on a vowel‐independent acoustic measure: the locus (virtual origin) of the second formant transition following the /d/‐release. Forty adult speakers of American English (20 male, 20 female) were recorded reading ten repetitions of 16 /dVd/‐words. From these data, individual F1‐F2 means and locus equations were calculated, including each speaker’s F2 locus frequency for /d/. Preliminary results based on the corner vowels /i, a, u/ show that F2 locus and F1‐F2 mean are strongly correlated (r‐squared=0.84). Statistical modeling is underway to determine whether scaling factors derived from locus analysis are superior to scaling factors derived from F1‐F2 means.