The Use of F0 Reliability Function for Prosodic Command Analysis on F0 Contour Generation Model

Abstract

This paper describes a method of utilizing an ``F0 Reliability Field'' (FRF), which we have proposed in our previous work, for estimating prosodic commands on F0 contour generation model. This FRF is the time-frequency representation of F0 likelihood, and an advantage of FRF is that it is not necessary to consider F0 errors that occur during an automatic F0 determination. Therefore, it is thought that FRF can be a more useful feature for automatic prosody analyses than F0 contour, and our previous paper has reported the validity of FRF on the analysis of detecting prosodic boundaries in Japanese continuous speech. Moreover, in this paper, we have examined the validity on the prosodic command estimation of superpositional model. Experimental results show that the accuracy of command estimation with FRF is well and it is close to the accuracy of command estimation with ideal F0 contour that has no F0 error.