ESSV Konferenz Elektronische Sprachsignalverarbeitung

Title: Evaluation of F0 Stylisation Methods and Fujisaki-Model Extractors

Authors: Hartmut R. Pfitzinger, Hansjörg Mixdorff


Four automatic methods for estimating parameters of the Fujisaki model are evaluated and compared with three F0 stylisation methods. Although the four methods yield comparable results with respect to their total errors, they show different error distributions. Particularly, the command amplitude distributions of two methods reveal weaknesses in accent or phrase command extraction due to arbitrarily set amplitude thresholds. Also, the means of the command rates are significantly different and their standard deviations are inhomogeneous. Finally, an alignment of the commands of the extractors shows correspondences between 46% and 91% of the phrase commands and 49% and 97% of the accent commands. In summary, two of the four Fujisaki-model extractors are currently unsuitable for meaningful phonetic as well as functional analysis and should be substantially improved.

Year: 2009
In session: Intonationsmodelle
Pages: 228 to 237