Investigation of coarticulation based on electromagnetic articulographic data

Coarticulation between the speech organs is a natural phenomenon in human speech. To realize this mechanism in our physiological articulatory model [1], this study attempts to analyze coarticulation involved in continuous speech based on electromagnetic articulographic (EMA) data recorded from three Japanese male subjects. The spatial target of each phoneme in CVC and VCV segments out of the sentences is determined using articulatory and acoustical cues, where V denotes five Japanese vowels, and C represents ten apical (including dental, alveolar, and postalveolar consonants) and two palatal consonants, respectively. A multiple regression analysis was applied to the extracted targets to evaluate "contributions" of the surrounding phonemes to the concerned target. It is found that the target of the central phoneme can be represented by a linear function of the preceding and following phonemes with a high accuracy. The result obtained form the segments with the apicals supports Oilman's conclusion derived from VCV utterances, that is, the articulation can be represented by a basic diphthongal gesture with an independent consonant gesture superimposed on its transitional portion [2]. The result from the segments with palatals suggested that the coarticulation caused by deformation was stronger than that induced by anticipation in the target planning stage.