Phonetic Detail in the Perception of Ethnic Varieties of US English 301are from voiceless ones in perceptual space. Overall, voicing is signaled less
well by four measures in the group with the most recent birth dates (0.351)
and the control group (0.399) as compared to the groups with older birth dates
(0.550, 0.479, 0.518, respectively). Table 13.3 reveals how important percent
glottal pulsing and vowel duration are to the data and the diminished statisti-
cal effect of a changing F1. Percent glottal pulsing is present in all groups,
and as the strongest measure in all four Wisconsin English groups. Moreover,
it never has a signi¿ cance value of more than 0.03. Vowel duration is present
in all groups with the exception of group 2. Change in F0 appears in three
groups (groups 1, 3, and 4) as the third or fourth strongest measure, never
reaching signi¿ cance below 0.10. Change in F1 appears somewhat strong in
group 1 (p<0.10), but inconsistently weak in group 3 (p=0.3640). Table 13.3
also yields an interesting observation that the measures selected from the
transformed perception data reÀ ects the selection of group 5 measures instead
of the group 3 or group 4 measures.
As mentioned previously, a common method of conducting a forward
stepwise discriminant analysis is to set a signi¿ cance threshold between 0.10
and 0.20. Such a threshold would allow a greater degree of strength to the
models for groups 2, 3, and 4. For group 2, in fact, only one measure, percent
glottal pulsing would model the data. The three other measures in the model
for group 2 add no more than 0.02 to the ASCC value. That is, adding the
three measures only accounts for 2% of the variation in the data. For groups
3 and 4, only percent glottal pulsing and vowel duration would make the sig-
ni¿ cance cut-off. The additional two measures account for an additional 1%
and 2% of the data variation, respectively.
Once the individual measures were selected for all groups using the for-
ward stepwise discriminant analysis, a canonical discriminant analysis was
conducted to verify whether a multivariate analysis is stronger than a univari-
ate one in terms of distinguishing voiced tokens from voiceless ones. Table
13.4 shows the univariate values for the four measures listed by descending
r^2 values for each group, while Table 13.5 shows relevant multivariate values
for each group. The squared canonical coef¿ cients (SCC) in Table 13.5 rep-
resent the amount of variation accounted for by the predictor variables as a
group. For the ¿ rst three test groups, the signi¿ cant variables in the model
(those whose inclusion is within the alpha limit) account for over 50% of the
variation (55%, 57% and 52%, respectively). In comparison, the amount of
variation accounted for by the signi¿ cant predictors is below 40% for the
fourth test group (35%) and the controls (40%). The model for the perceptual
group 3 data accounts for 46% of the variation. Comparing the SCC value
for each group in Table 13.5 with the best individual r^2 value in Table 13.4