Marcel Riedi
CHATR:
Modeling Prosody with MARS
and Text-to-Speech for German
Abstract:This technical report describes new models of segmental duration and fundamental frequency, and different improvements
and additions needed for German text-to-speech with CHATR.
The duration and frequency models were realized with "multivariate adaptive regression splines", a nonparametric regression method very well suited for problems having mixed ordinal
and categorical input factors. Models for German and English
have been developed. The text to-speech additions and improvements include modules for generating ToBI for German
text input and the handling of words not contained in the lexicon. "Decision trees" were used for these tasks.