Linear prediction coding of individual pitch accent shapes

The LPC-Intonation model presented here is dedicated to integrate speaker-dependent features into F0 modeling and analysis. It has been developed for an additional prosodic speaker-transformation based on the commandresponse approach for intonation features [1,2,3]. The basic principles of this integrated method for automatic analysis, coding, prediction and generation of F0 contours have already been presented in [7,9]. In this paper we are focusing on individual aspects of the resulting LPC filter parameters. Therefore we analyzed F0 contours of the 16 speakers in the German Phondat II Corpus. We found some individual differences of global parameters in speakers’ accent shapes that can be also modeled in synthesis. It can be shown that the resulting parameters are necessarily independent of linguistic information. Moreover, there are significant differences for each speaker. When transforming one speaker into another, a set of individual pitch tones for accents (’peaks’ and ’valleys’) and the resynthesis filter must be changed.