296 Thomas C. Purnell
to the gap duration (V:C ratio); the change in F1 from beginning of offset to
where it disappears (usually in the consonant); and the corresponding change
in F0.^10 We expect that vowel duration, percent glottal pulsing, duration of
glottal pulsing and the V:C ratio are larger for voiced tokens than voiceless
ones. F0, F1 and consonant duration were then transformed so that the voiced
values for all measures were larger than the measures for voiceless tokens.
Since the values for these three measures are linear, the transformation was
also linear using reference points to reverse the values.^11
The appropriateness of statistical tests applied to these measures focused
on the lack of truly independent variables and the identi¿ cation of underlying
latent factors, both of which are important in understanding the phenomena
of post-vocalic obstruent voicing. Regarding variable independence, many
acoustic characteristics of obstruent voicing overlap in non-experimentally
controlled tokens such that perhaps none of the purportedly independent vari-
ables are in fact independent, but rather co-vary with each other. Additionally,
a hint that at least one latent factor is operable in post-vocalic obstruent voic-
ing is found in the literature when references are made to a “low frequency
property” (Stevens and Blumstein 1981; Kingston and Diehl 1994, 1995).
This latent property is often thought to identify a combined percept involving
the acoustic characteristics of glottal pulsing, consonant duration, F1 and F0.
Thus, one prediction for an analysis looking for underlying factors would be
the emergence of a “low frequency property” aligned with the appropriate
acoustic characteristics. Tatsuoka (1970: 2) among others notes that overlap-
ping measures, such as the overlap of acoustic measures in the case at hand,
distorts a statistical relation across groups implied by univariate analyses.
As a demonstration of how inconsistently the acoustic measures co-vary
with each other, correlation matrices are shown in Tables 13.1 and 13.2 for
groups 2 and 3. Bilateral correlation comparisons of seven measures are
shown for both voiced tokens (the upper right triangle of data) and voiceless
tokens (the lower left triangle of data). Inconsistent co-variation is seen when
comparing which variables correlate with vowel duration for voiced tokens.
For group 3 voiced tokens (Table 13.2), the distribution of values of puls-
ing duration, percent glottal pulsing and the V:C ratio signi¿ cantly correlate
with the distribution of vowel duration. However, for group 2 (Table 13.1),
only V:C ratio signi¿ cantly correlates with vowel duration for voiced tokens.
Additionally, for both voiced and voiceless tokens, the percent of glottal puls-
ing (i.e., the pulsing duration divided by the consonant gap duration) is cor-
related with both the pulsing duration and consonant gap duration for both
groups of speakers. However, the inferential correlation, i.e., between pulsing
duration and consonant gap duration, is signi¿ cant for voiced tokens, but not
voiceless tokens.