Spectral Vowel Reduction in Japanese

1
Spectral Vowel Reduction in Japanese
May 2nd 2004

• Setsuko Shirai
• The University of Washington

2
Spectral Vowel Reduction

• Two kinds of spectral vowel reduction
• Phonological vowel reduction (lexical)
• "major" mejdZ?r
• majority m?dZ?r??i
• Phonetic vowel reduction (acoustic)
• I bought a book. ?
• Did you buy a book or many books? ej

3
Japanese reduction

• The Japanese language does not have a stress
  contrast so it is
• possible that no spectral vowel reduction occurs
  in Japanese.
• There is no known phonological vowel quality
  reduction in
• Japanese (as opposed to vowel devoicing).
• Question Does spectral vowel reduction occur in
  Japanese?

4
Previous study on formants
5
Hypothesis

• The Japanese function vowels, which are /a/,
  /e/, and /o/ following g_, d_ and t_
  respectively, will be spectrally reduced compared
  to their counter parts (i.e. content vowels).

6
Material

• Target vowels a, o and e
• ga subject marker
• to conjunction
• de indicates the place of action or copula
• 2 pairs of sentences for each target vowel
• (414 vowels)
• vowel /a/ 132 vowels
• vowel /e/ 138 vowels
• vowel /o/ 144 vowels

7
Tokens

• Content Function
• jinbutsuga junbutsu-ga
• portrait people subject marker
• kiga ki-ga
• starvation tactful
• hato ha-to
• pigeon leaf - conjunction

8
Token (cont.)

• Content Function
• kogoto kogo-to
• scolding old-Japanese conjunction
• tade ta-de
• smartweed rice field - at
• hitode hito-de
• starfish person - copula

9
Recordings

• 12 Speakers 6 males 6 females Tokyo dialect
  speakers
• Most of subject 20s ESL students
• Place the sound attenuated recording booth in
  the phonetic lab at UW
• Reading 5 randomized sentence lists, of which 3
  were used
• the first and the last readings not used

10
Recordings (cont.)

• Recording
• Electro-Voice RE20 microphone with a flat
  response
• to 20 KHz Analog cassette tape recorder
• (TASCAM 122 MKIII)
• Digitizing
• Sound Edit 16 version 2 on a computer
• with an Audiomedia III card
• Sampling rate 11025 Hz, 16 bit

11
Measurements

• F1 and F2 are measured at 5 points
• at the beginning, 1/4, 1/2, 3/4, and at the end
  of each vowel
• ? formants
• ? formants F1 at the beginning - F1 at the
  middle
• ? formants are an indicator of formant
  movements.
• For /a/ F1 difference is relevant for measuring
  reduction
• For /e/ and /o/ F2 difference is relevant for
  measuring reduction

12
Figure 1 Formant plot
Content red, Function blue
13
Table 1 mean SD of F1
Almost 90 Hz difference in F1 /a/ - indicates
vowel reduction.
14
Table 2 mean SD of F2
20Hz and 15Hz differences in F2 of /e/ and /o/
respectively. The differences are subtle, but
indicate vowel reduction.
15
Figure 2 box plot illustrating F1
Relatively large difference in F1 for /a/, but
not for /e/ or /o/.
16
Figure 3 box plot illustrating F2
No large differences (content/function) for /e/
and /o/.
17
Results of a repeated measure ANOVA

• There is a significant LEXICAL effect on F1 of
  /a/no significant LEXICAL effects on F2 of /e/
  or /o/
• on F1 of /a/ ( F1, 65 73.401, p
• Post Hoc Bonferroni test with ? 0.05 -
  significant
• on F2 of /e/ (F1, 68 3.549, p 0.067)Post
  Hoc Bonferroni test - not significant
• on F2 of /o/ (F1, 71 1.150 p 0.287)
• Post Hoc Bonferroni test - not significant

18
Discussion Reduction in Japanese?

• Japanese function vowels are centralized
• F1 of function /a/ lower than F1 of content /a/
  indicating vowel raising
• F2 of function /e/ lower than that of content /e/
  indicating vowel backing
• F2 of function /o/ higher than that of content
  /o/ indicating vowel fronting

19
Discussion (cont)Reduction in Japanese?

• Function /a/ shows statistically reliable
  spectral reduction, but other vowels do not.
• Possible reason
• Locus target distance between vowels and /d/ or
  /t/ are short ? the magnitude of lexical effect
  small

20
Discussion (cont)Reduction in Japanese?

• Japanese function vowels are centralized.
• Japanese function vowels are shorter than content
  vowels (Shirai 2002).
• Next question Is duration sole determinant of F1
  displacement in /a/?

21
Hypothesis 2

• If undershoot is the source of reduction, then
  differences in F1 will be linked to differences
  in duration.
• If centralization is reduction that is
  independent of undershoot, then differences in F1
  will be observed in the absence of differences in
  duration.

22
Trajectories of F1 of /a/

• To test whether or not the duration is sole
  determinant of formant displacement, trajectories
  of F1 are used.
• If undershoot is the explanation, then the F1 of
  function /a/ should move towards the target of
  the F1 of content /a/.

23
Figure 4 Trajectories of F1 of /a/
target
onset
Onset F1 of function /a/ (blue) does not coincide
with onset F1 of content /a/ (red), which
replicates van Bergems (1993) result for Dutch.
24
Categorical duration

• To test whether or not there is lexical
  difference when the duration is the same, the
  duration is classified into 10 categories.
• If undershoot is responsible for F1
  centralization in function /a/, then different F1
  values should be restricted to different
  durations.
• If centralization is caused by reduction, then
  different F1 values should be seen at the same
  durations.

25
Figure 5 Scatter graph illustrating F1 of /a/
F1s of content /a/ (red) are higher than F1s of
function /a/ (blue) where both durations are the
same.
26
The results of a factorial ANOVA with F1
categorical duration

• A factorial ANOVA with F1 is as dependent
  variable, LEXICAL and CATEGORICAL DURATION as
  independent variables
• ResultsLEXICAL F1, 112 4.072, p 0.046
  DURATION F9, 1121.869, p 0.064
  INTERACTION F9, 112 0.909, p 0.520
• There is a significant effect for Lexical status
  (content/function) for F1, but no significant
  effect for duration, nor a significant
  interaction between the two. This means that
  while function /a/ is more reduced than content
  /a/, the reduction is not tied to duration.

27
Trajectories of ? F1 for /a/

• If centralization is related to reduction more
  than undershoot, we expect there to be a
  different target for function /a/ from that of
  content /a/.
• To test these hypotheses, the formant
  trajectories were converted to ? F1 so that they
  both have the same starting point.

28
Figure 6 ? F1 by duration
target
? F1 of function /a/ has a different target than
content /a/.
29
? F1 and Categorical duration
Again, the duration is classified into 10
categories. If undershoot is responsible for F1
centralization in function /a/, then different ?
F1 values should be restricted to different
durations. If centralization is caused by
reduction, then different ? F1 values should be
seen at the same durations.
30
Figure 7 Scatter graph illustrating ? F1
? F1 of content /a/ is different than ? F1 of
function /a/ even when the duration of function
/a/ is long.
31
Results of Factorial ANOVA with ? F1
categorical duration

• A factorial ANOVA with ? F1 as a dependent
  variable, and CATEGORICACAL DURATION and LEXICAL
  as independent variables.
• ResultsLEXICAL F1, 112 6.814, p
  0.010DURATION F9, 112 4.860, p 0.001INTERACTION F9, 112 0.571, p 0.818
• There is a significant LEXICAL effect on ? F1 of
  /a/ and for DURATION effect, but there is no
  significant INTERACTION between the two. Again
  this means that centralization is not tied to
  duration.

32
DiscussionUndershoot or centralization?

• F1 of function /a/ is lower (i.e. more
  centralized) than that of content /a/ even when
  the duration is long.
• The onset F1 of content /a/ does not agree with
  the onset F1 of function /a/ replicating van
  Bergems findings for Dutch.
• The results of ANOVA the lack of interaction,
  no significant duration effect, a significant
  lexical effect indicate that observed
  centralization is independent of the observed
  differences in duration.

33
Discussion (cont)Undershoot or centralization?

• ?F1 of function /a/ is lower than that of content
  /a/ even when the duration is long.
• The results of ANOVA
• a significant lexical effect,
• a significant duration effect,
• the lack of interaction
• indicate that the trend line for ? F1 of
  function /a/ is parallel to the trend line for
  that of content /a/.
• The target for function /a/ is different than
  that for content /a/.

34
Conclusion

• Spectral vowel reduction occurs in Japanese.
• Duration is NOT sole determinant of formant
  displacement.
• Spectral vowel reduction is the combination of
  the result of short duration, contextual
  assimilation and difference of the targets.

35
Acknowledgement

• I would like to thank the subjects who
  participated in my research.
• I appreciate the valuable advice from Richard
  Wright, Alicia Beckford Wassink, Sharon Hargus,
  and other members of the UW Phonetics Lab.
• And, last but not least, I would like to thank
  you for listening.

36
REFERENCES

• Keating, P., and Huffman, M. 1984. Vowel
  variation in Japanese. Phonetica, 41, 191-207.
• Lindblom, B. 1963. Spectrographic Study of Vowel
  Reduction. Journal of Acoustical Society of
  America, 35, 1773-1781.
• Moon, S. J., and Lindblom, M. 1994. Interaction
  between duration, context, and speaking style in
  English stressed vowels. Journal of Acoustical
  Society of America, 96, 40-55.
• Nearey, T. M. 1978. Vowel-space normalization
  procedure and phone-preserving transformations of
  synthetic vowels. Journal of Acoustical Society
  of America, 63, S5.

37
REFERENCES

• Nord, L. 1986. Acoustic Studies of Vowel
  Reduction in Swedish. Quarterly Progress and
  Status Report, 4, 19-36.
• Shirai, S. 2002. The duration of Function Words
  in Japanese. UW Working Papers in Linguistics.
  21.
• Van Bergem, D. R. 1993. Acoustic vowel
  reduction as a function of sentence accent, word
  stress, and word class. Speech Communication,
  12, 1-23.
• Wright, R. 2003. Lexical Competition and
  Reduction. Papers in Laboratory Phonology VI.