Targetless /u/ in Tokyo Japanese
Shigeto Kawahara 1 , Jason A. Shaw 1,2,3 , James Whang 1,4
Keio University
1, Western Sydney University
2, Yale University
3, New York University
4In Tokyo Japanese, high vowels tend to devoice
between voiceless consonants:
Research Questions:
Do devoiced vowels have lingual targets?
Are lingual gestures reduced when inaudible?
High vowel devoicing Electromagnetic Articulography (EMA)
Determining target specification: DCT, Micro-prosodic sampling, Bayesian Classification
Results
Equipment: NDI Wave EMA system
used to tranduce movements of three
fleshpoints on the tongue:
Participants: 6 (3 male) Tokyo natives aged 19-22
Materials: 5 near minimal pairs with voiced/voiceless
/u/ read in carrier phrase: ookee ___ to itte ‘okay, say
___’; 10-15 repetitions per word.
Tongue Dorsum (TD)
ALL OR NOTHING
• Devoiced vowels are variably targetless; the probability
of targetlessness:
(1) varies across speakers
(2) is conditioned by linguistic factors
• Phonetic reduction, on the other hand, is uncommon.
ɸu̥soku
~
ɸuzokuʃu̥taise:
~
ʃudaikakatsu̥toki
~
katsudouhaku̥sai
~
yakuzaimasu̥taa
~
masudaDiscrete Cosine Transform (DCT) to reduce
dimensionality of the data
Define stochastic generators of target and
targetless hypothesis
Bayesian Classification parameters (for ʃu̥taise: )
Possible outcomes expressed as posterior probabilities
Compute Bayesian posterior probabilities
of targetlessness
S01 S02 S03 S04 S05 S06
Posterior probability
of targetlessness
TD vertical position over time (VCuCV sequences)
0.47 0.39 0.75 0.84 0.01 0.19
0.92 0.68 0.84 0.99 0.02 0.89
0.44
0.72
0.56
0.18
0.48
• Japanese has consonant clusters (sometimes) in that the TD
follows a linear trajectory from C1 to C2 for many tokens.
• Syllable contact laws may influence the appearance of clusters:
• Results support a phonological view of high vowel deletion over
gestural overlap
(e.g., Kondo, 2001 vs. Jun & Beckman, 1993).• The computational tools can be broadly applied to detect
phonological (under-)specification in phonetic data.
Acknowledgments: Supported by JSPS fellow ship to Jason A. Shaw (#15F15715) and grant-in-aid (#26284059) to S. Kawahara; thanks to Chika Takahashi and Jeff Moore for help with data. 0.81 0.19 0.69 0.93 0.06 0.79
0.00 0.00 0.51 0.50 0.00 0.07
0.64 0.09 1.00 0.02 0.74 0.41