Phonetic Detail in the Perception of Ethnic Varieties of US English 313ethnically af¿ liated speech community. The speci¿ c case at hand involves a
marker of certain European immigrant communities (¿ nal devoicing). The
¿ rst hypothesis we tested was whether there was a statistically signi¿ cant dif-
ference between Wisconsin English speakers (those who may have integrated
this low-level ethnically-af¿ liated cue) and the American English controls
(who we assume from experience do not use devoicing in the same way). We
found that all groups share the importance of percent glottal pulsing in signal-
ing the voicing distinction, and most Wisconsin English groups share vowel
duration as an important measure of voicing. However, while additional mea-
sures contribute to the models, the set of important measures is not uniform
across all groups. Additionally, the amount of variation the set of measures
accounts for and the individual contribution to the model varies by the vary-
ing eigenvector weights. The second hypothesis we tested with this data was
whether an analysis of the acoustic data and an analysis of the acoustic data in
light of perception results are identical. Thus the relation depicted by Figure
13.1(d) is important. The group 3 perception data showed an interesting pat-
tern in which the measures which appear important are those of group 5, yet
the weighting—via the identi¿ cation of latent factors—appears aligned to a
Wisconsin English group (group 2). It is unclear from this data how ages of
speaker and listener interacts with weighting of acoustic cues as they are used
in perception. However, ¿ ndings of the principal component analysis suggest
that Wisconsin English speakers use vowel duration and percent glottal puls-
ing in a trading relation. These ¿ ndings conÀ ict with Kingston and Diehl’s
claim that speakers of American English use enhancement.
Three issues make the study of Wisconsin English devoicing particularly
interesting. First, the variables are low-level phonetic attributes and not seg-
mental ones. Second, this study explores variation in consonants in a way
that is largely overlooked by sociolinguists.^14 Third, the low-level phonetic
cues are tied to the sub-segmental distinctive features in a robust statistical
fashion. The canonical discriminant analysis can reveal only one latent factor
across the entire data, VOICING, assuming binary features. We now turn to a
case where low-level variation will be examined across three ethnically af¿ li-
ated groups, thereby opening the possibility for more than one latent factor.
- Example 2: “hello” in three dialects
4.1 Background
The second set of data which shows variation in the mapping of acoustic char-
acteristics to perceptual cues comes from experiments using a tri-dialectal