Processing of Speech Variability: Vowel Reduction in Japanese

Abstract

This dissertation investigates the processing of speech variability, allophonic and indexical variation in Japanese. A series of speech perception experiments were conducted with reduced and fully voiced vowels in Japanese as a test case. Reduced vowels should be difficult for listeners to hear because they are acoustically less salient than fully voiced vowels, due to the lack of relevant physiological properties. On the other hand, reduced vowels between voiceless consonants represent more common phonological patterns than fully voiced vowels. Furthermore, previous studies found that Japanese listeners were capable of hearing completely deleted vowels. Listeners intuitively maintain CV syllables in perception, hearing a vowel after each consonant in order to avoid obstruent clusters (a violation of Japanese phonotactics).It was found that listeners made good use of acoustic, phonological, and phonotactic knowledge of their native language for processing allophonic variants. In word recognition, listeners performed better when reduced vowels were in the environment where vowel reduction was expected. The phonological appropriateness of an allophone was judged in relation to adjacent consonants on both sides, and the facilitatory effect of appropriateness of reduced vowels surpassed the inhibitory effect of their acoustic weakness. However, in terms of sound detection, listeners found reduced and fully voiced vowels equally easy to hear in an environment where vowel reduction was expected. Although reduced vowels were phonologically appropriate between voiceless consonants, the phonological appropriateness merely balanced out acoustic weakness; it was not strong enough to surpass it. In addition, the phonological appropriateness of an allophone was judged based only on the preceding consonant, which suggests that listeners processed sounds linearly. Furthermore, the study found that phonological appropriateness of the allophone was affected by dialectal differences and speech rates. Listeners' preference for a certain allophone was influenced by the phonology of a listeners' native dialect and expectation was skewed by fast speech rates.This study suggests that current speech perception models need modification to account for the processing of speech variability taking language-specific phonological knowledge into consideration. The study demonstrated that it is important to investigate at which stage phonological inference takes place during processing.