Detecting accent sandhi in Japanese using a superpositional F0 model

In this report, we propose a method for automatic prosodic structure recognition of Japanese utterances based on a superpositional F0 model, focusing particularly on the accent sandhi phonemenon in compound nouns. The method enables automatic labeling of F0 contours using the model, which can be useful for creating prosodic databases containing F0 contours in a parametric form. The prosodic structure is identified by comparing the distances between F0 contours generated by hypothetical model configurations and the extracted F0 contour, and choosing the configuration that yields the smallest distance. In this paper, we apply the method to detect the accent sandhi pattern of compound nouns made up of 2 or more words, and show that the method can correctly identify their prosodic structure, except for 1-mora deviations in the position of the accent nucleus.