Abstract
A method of representing Japanese vowels is proposed based on the Principal Component Analysis of the estimated vocal tract area functions. The vocal tract area functions are estimated by an adaptive inverse filtering method and normalized assuming that the volume of the vocal tract is constant. The experimental results are summarized as follows. In case the analysis is limited to single vowels from a group of male speakers, the cumulative contribution rate for the first 3 vectors reaches 92. 3%. in this dimensionality each axis corresponds to the articulatory structures of the vowels and the separation among vowel groups is completed. Then the variations of the area functions of single vowels can be represented by a linear model of 3-dimensions. The linearity of the variations holds also for the vowels uttered continuously. When a difference in the vocal tract length between males and females is taken into account, the cumulative contribution rate of vocal tract variations reaches 91. 2% in the first four principal axis, and the separation of vowels is completed also in this dimensionality. When the child's vocal tract functions are considered together with those adult materials, the dimensionality of the variations is the same as those in the case of the group of males and females. Then the variations of the vocal tract functions with extreme differences in the vocal tract length can be represented by a linear model of 4-dimensions.
