Normalizing pitch variability in intoxicated speech. Ryan Hanke

1
Normalizing pitch variability in intoxicated
speech.Ryan Hanke Thomas PurnellDepartment of
LinguisticsUniversity of Wisconsin-Madison
3aSC25
Table 3. Results of Experiment 3.

• BACKGROUND
• Previous Studies Findings
• F0 rise under intoxication (Behne and Rivera
  1990, Hollien et al. 2001)
• F0 lowering under intoxication (Johnson et
  al..1990, Watanabe et al. 1994, Alderman et al.
  1995)
• F0 variability increase (Pisoni et al. 1985,
  Klingholz et al. 1988, Behne and Rivera 1990)
• Purpose of Preliminary Study
• the effects of intoxication on measures of
  fundamental frequency
• the extent to which such variables as phrase
  size and subjects influence intoxicated speech
  pitch studies
• the effect of placebo on pitch in intoxicated
  speech
• These conflicting findings may be due to
  increased F0 fluctuations at various levels of
  intoxication, and use of inconsistent prosodic
  domains, such as averaging over entire sentences
  or utterances. If phonetic analysis takes into
  account prosodic boundaries, then models will
  account for appropriate prosodic rises and falls
  (Taylor 2000).
• Three Hypotheses
• Hypothesis A Under intoxication, pitch rises
  from a subjects mean
• Hypothesis B Under intoxication, pitch falls
  from a subjects mean
• Hypothesis C Under intoxication, pitch varies
  from a subjects mean
• EXPERIMENT 1
• Goal

Figure 1. (top) Individual subject data for mean
of the change in mean F0 minus the baseline.
(bottom) Multidimensional scaling of normalized
(semitones, rotated) measures. Note the change to
maximum and minimum F0 as well as slope.
EXPERIMENT 2 Goal To test Hypotheses A and B in a
placebo group. Method The 4 male and 2 female
placebo condition subjects received fruit juice
with what they thought was vodka, water in a
vodka bottle. The procedures and reading
materials were the same as the alcohol and
control conditions in Experiment 1. Results
Discussion Significance of the main effects in
the placebo condition (Table 2) indicates
increasing variability in behavior with
increasing levels. This variability, not seen in
the control condition, demonstrates the placebo
effect the subjects behavior suggests that
thought they were becoming intoxicated when in
fact they were not. The placebo conditions
results parallel those of the alcohol
conditions.
EXPERIMENT 3 Goal To test Hypothesis
C. Method The subjects, tasks, and conditions
remain the same as in Experiments 1 and 2. Each
of the 7 phrases used per level were analyzed.
Using individual phrases allows for the control
of prosodic patterns instead of averaging over
entire utterances. The pitch stylizations linear
contour was rotated by subtracting linear fit. F0
standard deviation, minimum, maximum, and range
were then calculated. Mean and standard deviation
in semitones (Henton 1989, Viscovich et al. 2003)
and original F0 slope were also included in the
analysis. Results Discussion For the alcohol
and control conditions, none of the variables
show a significance for the Level main effect. A
number of normalized variables show a
significance for the Subject main effect for both
conditions (Table 3). Semitones standard
deviation shows a significance Level main effect
(F(2, 115)3.135, p0.047). The variables in
Table 3 show a significance for the Subject main
effect. Given the number of these variables that
show significance, it appears that more than just
the traditional pitch measures, F0 mean and
standard deviation, are important cues in
intoxicated speech. To determine which of these
normalized variables are most important, a
multidimensional scaling analysis was performed
on these variables for each of the 3 groups
(stress0.0114, RSQ0.9995). The result that
emerges is that maximum F0, minimum F0, and slope
are the leading cues. They show similar patterns
with respect to one another in the control and
placebo conditions. In the alcohol condition,
however, maximum F0 lowers on dimension 2 and
minimum F0 and slope trade places. As these
variables are normalized, remaining differences
are a function of subjects individual
differences. The largest number of variables
showing significance occurs in the placebo
condition, reaffirming the placebo effect. It may
be tempting to conclude, then, that the placebo
condition shows the highest degree of
variability. However, the multidimensional
scaling solution shows a pattern in the alcohol
condition that is different from the control and
placebo conditions. It may be the case that the
overall pattern and how variables interact with
each other is a better measure of variability.
This tentatively supports Hypothesis C above.
Table 1. Results of Experiment 1.
DISCUSSION Subjects in the placebo condition
showed an increase in mean F0, like the subjects
in the alcohol condition, but showed a pattern in
the multidimensional scaling solution much like
that of the control condition with respect to
slope, minimum and maximum F0. This finding fits
with that of Hollien et al. (2001). They found
that F0 is one cue actors manipulate when sober
attempting to sound intoxicated, and when
intoxicated attempting to sound sober. Some cues
appearing in intoxicated speech those that
speakers can manipulate when producing artificial
intoxicated speechare more salient than others.
The lower level cues, though, may be unique to
truly intoxicated speech. CONCLUSION Preliminary
Research while pitch does rise (supporting
Hypothesis A), the variation by subject
influences the output there is greater
variability of some cues in the alcohol condition
(supporting Hypothesis C) Future Research more
subjects examine between subject variation
multivariate analysis, such as confirmatory
factor analysis, to differentiate salient cues
from lower level cues Contact Rya
n Hanke Thomas Purnell rmhanke_at_wisc.edu tcpurn
ell_at_wisc.edu
Table 2. Results of Experiment 2.