316 Thomas C. Purnell
Use of these seemingly disparate cues by listeners is consistent with stream
segregation and perceptual grouping associated with other processes in audi-
tion-based cognition (Bregman 1990).^19 Univariate results are shown on Table
13.9 and multivariate results appear on Tables 13.10 and 13.11. Unlike in the
analysis on obstruent voicing, we see two latent factors emerge in Table 13.10.
Although the univariate r^2 for /ܭ/ backness is fairly high, (0.655, Table 13.9), the
multivariate model increases the r^2 value by over 20% (to 0.870, Table 13.10).
For the perception data it is clear that the ¿ rst factor r^2 (0.768, Table 13.10) is
much higher than the best univariate r^2 (0.455, Table 13.9).
The principal component analysis used to identify within-group varia-
tion indicates that for some measures such as those used in the canonical
discriminant analysis, three latent factors emerge for both acoustic and per-
ceptual data. Following the procedure from §3, the strongest value across all
Table 13.8 Results of Stepwise Discriminant Analysis Using a Forward Stepwise
Method for AAE, ChE and SAE “hello” Tokens (Į= 0.10 to enter model)Va r iable Partial r^2 F p > FAve r age
squared
canonical
correlation
Acoustics (df = 12, 46)
/ܭ/ F3-F2 (Z) 0.655 25.609 0.0001 0.327*
/o/ F2 (Z) 0.584 18.259 0.0001 0.588*
RMS Standard Deviation 0.405 8.507 0.0015 0.681*
RMS Initial Value 0.312 5.450 0.0112 0.722*
Maximum Flow Declination Rate 0.341 4.942 0.0083 0.791*
/ܭ/ F1 (Z) 0.199 2.732 0.0872 0.816*
Perception (df = 10, 46)
/o/ F2 (Hz) 0.455 11.255 0.0003 0.227*
/ܭ/ F3-F2 (Z) 0.378 7.888 0.0021 0.378*
Change in RMS 0.343 6.523 0.0053 0.506*
/ܭ/ F1 (Z) 0.245 3.889 0.0344 0.570*
/o/ F3-F1 (Z) 0.273 4.328 0.0254 0.620*
Note: Each dialect had ten representative tokens.
*p<0.001