Perceptieeffecten van geografische variantie in microprosodische eigenschappen

1
Perceptie-effecten van geografische variantie in
(micro-)prosodische eigenschappen

• Marie Nilsenová
• Communicatie- en informatie Wetenschappen
• Faculteit Geesteswetenschappen
• Universiteit van Tilburg

2
Background

• Clear correlation between age, body size and F0,
  as well as timbre (Greisbach 1999, Kraus,
  Freyberg Morsella 2002)
• Listeners should be able to exploit that
  knowledge, eg reflecting on the Frequency Code
  deeper lt-gt bigger/taller
• But listener estimates not as good as could be
  expected, also correlation not always perfect
  (Cohen et al 1980, Gunter Manning 1982, van
  Dommelen Moxness 1995)
• Free (non-linguistic) variation with social,
  geographical distribution?
• Expectation listeners should perform better when
  estimating speakers in their own variant compared
  to another variant

3
Methodology

• 16 speakers (8 male, 8 female/8 B-Dutch, 8
  B-Flemish) reading the same short text (newspaper
  article)
• 2 groups of participants - Dutch, Flemish
• Task estimate speakers length, weight, age and
  the pleasantness of the voice (operationalized
  as 3 properties Cronbachs ? gt.65) fixed time
• Also report if speaker Dutch or Flemish

4
Speaker comparison Female
5
Speaker comparison Male
6
Questionnaire

• Wat vindt u van deze stem?
• Plezierig ? ? ? ? ? Onplezierig
• Vriendelijk ? ? ? ? ? Onvriendelijk
• Aangenaam ? ? ? ? ? Vervelend
• Wat is volgens u de leeftijd van de spreker?
  jaar
• Wat is volgens u de lengte van de spreker?
  cm
• Wat is volgens u het gewicht van de spreker?
  kg
• De spreker is Vlaams ? Nederlands ?

7
Design

• Two conditions normal and reversed speech
  (Audacity)
• Purpose of the reversed speech condition
• to assure that speakers do not adapt to the
  variant they hear based on previous experience
• also, to examine if adaptation to own variant
  also relevant for perception of an unknown
  language

8
Participants (N179)
9
Measure of accuracy

• Index of accuracy (Kraus, Freyberg Morsella
  2002) the average of the absolute difference
  between estimated and actual values (AD) - the
  mean of the absolute differences between judges
  estimates of speakers values on an attribute and
  the speakers actual value
• The AD measure indexes judges average error in
  estimating a particular attribute
• It answers the question How close is the average
  estimate of attribute X to the actual value of
  X?
• Absolute accuracy - the sum of the absolute
  difference between estimated and actual values
  (AA)

10
Results (Length)

• Normal condition (AD) interaction effect between
  speakers variant and gender, and listeners
  variant (plt.05, partial eta squared .127) in
  the predicted direction
• Reversed condition (AD) main effect of speakers
  variant (plt.05, partial eta squared .058)
  Dutch speakers estimated more accurately than
  Flemish speakers

11
(No Transcript)
12
(No Transcript)
13
Results (Pleasantness)

• Normal speech condition interaction effect of
  speakers variant and listeners variant (plt.01,
  partial eta squared .135) - listeners prefered
  their own variant
• Reversed speech condition clear main effect of
  speakers variant (plt.001, partial eta squared
  .535) - reversed Flemish prefered to reversed
  Dutch

14
Conclusion

• Accuracy of length estimation higher in normal
  speech condition for own variant
• Result not translated into the reversed speech
  condition
• Clear preference for speakers of own variant
• Listeners generally reported that estimating
  weight difficult
• Other measures of accuracy to explore the
  directions of the accuracy judgments

15
Thanks to

• Hanne Kloots
• Machiel Debets
• Julia Hirschberg
• Marc Swerts