Robust unit selection based on syllable prosody parameters

In this paper we present a robust unit selection based on syllable prosody parameters (RUSSPP). After almost isolated neural network (NN) predictions of f0-contours and segmental durations these parameters are re-utilized for a search in our database for best fitting speech elements and acoustic prosody parameters. This search is realized by using a Viterbi-algorithm that operates on syllable level but explicitly allows higher and lower level speech elements. Dealing with a limited speech corpus for synthesis makes signal processing necessary for speech elements at concatenation points. Also a simple but efficient post-processing on the selected prosody parameters is introduced. This new method is applied and tested within our TTS system PA-PAGENO for a German male news speaker. It could be shown that it improves the quality of our prosody generation module and of the selection process.