226 Bartáomiej Plichta and Brad Rakerd
2.1 Intrinsic normalization
Proponents of intrinsic normalization theory argue that there is suf¿ cient
information within the spectral content of the vowel itself to support a listen-
er’s normalization computation. Two of the most convincing studies of intrin-
sic normalization are those by Miller (1989) and Syrdal and Gopal (1986).
Miller demonstrated that the monophthongal vowels of American English can
be represented as distinct clusters or “target zones” within a three-dimen-
sional perceptual space, where the dimension correspond to: (i) the difference
between the center frequencies of formants three and two (F3-F2); (ii) the dif-
ference between formants two and one (F2-F1), and the difference between
formant one and a sensory reference (F1-SR). Miller examined a number of
speech corpora in this way, and found that essential vowel category contrasts
can be accurately represented by means of this model.
Syrdal and Gopal used a Bark transformation to devise a two-dimensional
model of vowel recognition based primarily on the perception of critical dis-
tance, in Bark (Chistovich 1985). Their evidence for intrinsic normalization
came from successful discriminant analysis of vowels. The discriminant
plane resulting from the analysis was delimited by parameters derived from
the ratios of Bark-transformed F1-F0, F2-F1, and F3-F2.
A limitation of these studies is that they did not expressly include substan-
tial sources of sociolinguistic variability as it relates to vowels. Both studies
used the Peterson and Barney corpus, which contains a well-documented but
limited database of formant values from a select set of monosyllabic words,
all produced in /hVd/ context. Very little is known about the speakers’ dialect
history, however, and no direct attempt was made to capture dialectal varia-
tion within the corpus. Other test corpora were similarly lacking in docu-
mented sociolinguistic variability.
To illustrate the potentially important role of sociolinguistic variation, a
Bark analysis similar to that of Syrdal and Gopal (1986) was applied here to a
cor pus of monophthongal vowels elicited f rom 26 NCCS speakers and, for com-
parison, to the vowels of the Peterson and Barney corpus. Figure 9.3(a) shows
a two-dimensional plot of the Peterson and Barney data delimited by F3-F 2
and F1-F0 in the discriminant plane. It can be seen that each vowel occupies
its own unique spot in the discriminant space (delimited by the ellipses), which
suggests that with intrinsic normalization there should be very little vowel con-
fusion. Results for the NCCS vowel set are quite different. There is substantial
overlap among /a/, /͑/, /ࣜ/, a n d /ɬ/ in this corpus, which suggests that vowel con-
fusion would be more likely to occur in communities where NCCS is in prog-
ress (Preston, this volume). These results also suggest that in such communities
the success of intrinsic normalization strategies is less likely.