Mastering Japanese pitch accent is a significant hurdle for learners. To effectively support accent acquisition, quantitative evaluation of pitch accent is crucial. Our research re-examines the properties of pitch accent in four-mora words, shifting the focus from how speakers produce it to how listeners perceive it. We utilized a Sequential Variational Autoencoder model to extract feature representations of pitch accent in four-mora Japanese words. Then, we generated spoken words with various F0 contours, each with distinct accentual features. Native Japanese speakers rated the appropriateness of these generated words. Our findings showed that even words with the same accent type in speech corpora had different appropriateness distributions depending on their lexical properties. This presentation further explores how listener and speaker metadata—specifically, speaker gender, listener gender, and listener dialect—influence these appropriateness rating distributions. We used the two-dimensional Kolmogorov–Smirnov test to identify significant differences across these metadata groups. While we didn't find significant differences for listener gender, we observed several significant distinctions related to speaker gender. For dialect, we compared Tokyo and Keihan styles, noting significant differences for certain lexical properties. Furthermore, by applying a Gaussian Mixture Model to the rating distributions, we discuss which accent types were highly rated.
Masuda-Katsuse et al. (Wed,) studied this question.