Abstract
Large-scale corpus of spontaneous speech can be an excellent resource for the study of language variations. This philosophy was reflected in the architecture of the Corpus of Spontaneous Japanese, or CSJ, that we have been compiling since 1999. In the first half of the paper, various characteristics of the CSJ were described in view of its application for the study of language variations. In the last half of the paper, the word-fusion of "de+wa" into "zya" was analyzed using the CSJ. It turned out that the main factor of the word fusion involved part-of-speech of "de" (either particle of auxiliary verb), type of speech (academic presentation or public speech), the degree of speaker's relaxedness, as measured by the frequency of laughter, and the interaction between the speech type and the frequency of laughter. The analyses of the interaction term suggested that low-frequency variations could be better perceptual cues of style shifting than high-frequency variations.
