Distinctive features in a Hungarian hidden Markov model based TTS system

This paper describes the improvement of Hungarian hidden Markov model (HMM) speech synthesis by introducing distinctive features. The values of 18 distinctive features for the Hungarian phonemes are determined and applied to an existing HMM speech synthesis system: they are used in context-dependent labeling and in the decision tree structure. The results show that some distinctive features have become significant nodes of the decision trees. The improvements are evaluated by the analysis of the decision trees and by a listening test as well.