312 Thomas C. Purnell
The results from this type of study can also be used to distinguish an anal-
ysis, which is based on trading relations or feature enhancement. If two mea-
sures are in a trading relation, then we can see that when one measure is weak
for some category member (e.g., the percent glottal pulsing is below 50% of
the consonant for voiced tokens), then an opposing measure is strong (e.g., the
vowel duration would be longer, compensating for the low pulsing). Thus, we
expect that these two values are complementary and either display a signi¿ cant
negative correlation or display no signi¿ cant correlation. In contrast, if two
measures are in an enhancement relation, then we can see that both features
are simultaneously strong, or display a signi¿ cant positive correlation. In the
papers by Purnell et al. (2005a, b), the argument is made for a trading rela-
tion between vowel duration and percent glottal pulsing for Wisconsin English
post-vocalic obstruent voicing. This differs from the claim in work by Kingston
and Diehl that measures of post-vocalic obstruent voicing are enhancing. Cor-
relations were calculated by voicing category (voiced, voiceless) and by data
group for the four measures identi¿ ed as strongest for that group. The correla-
tion between vowel duration and glottal pulsing was examined for all groups
except group 2 which did not identify vowel duration as a signi¿ cantly strong
measure. The two measures were signi¿ cantly correlated for some, but not all
instances. Group 1 voiceless tokens show a signi¿ cant negative correlation
(- 0.5330, p<0.05), but not the voiced tokens (0.3176, p >0.05). Conversely, for
the group 3 perception data the voiced tokens show a signi¿ cant negative corre-
lation (-0.4529, p <0.05) while the voiceless tokens do not (-0.1795, p =0.3340).
For group 3 and group 5 the voiceless tokens (-0.3331, p <0.01; -0.3019, p <0.05)
and the voiced tokens (-0.2846, p <0.05; -0.4411, p <0.01) show that vowel
duration and percent glottal pulsing are negatively signi¿ cant. Neither group
4 voiceless or voiced tokens are signi¿ cant for the two measures (0.0160, p
=0.9410 and 0.0626, p =0.6938, respectively). Thus, when there is a signi¿ cant
correlation, it is only negative. This tentatively argues in favor of a trading rela-
tion model rather than an enhancement one for all cases that use vowel duration
and percent glottal pulsing, including group 5 controls. Additionally, Purnell
et al. (2005b) claimed that the group 4 data was interesting because the voiced
tokens had moved very close to the voiceless tokens (paralleling ¿ nal devoic-
ing of a native German control). The closeness of voiced and voiceless tokens
might account for the lack of signi¿ cant correlation.
3.4 Conclusions
The larger task at hand is to describe variation in the phonetic compo-
nent through which a speaker is perceived by a listener as belonging to an