Text-to-speech synthesis based on grapheme and phoneme clusters

We are developing a text to speech synthesis system for the German language, which has the sametype of speech elements in both, the linguistic-phonetic transcription and the phonemization Ievei, namely 'clusters'. They are defined as sequences of graphemes or phonemes of the same type (vowel or consonant clusters). Compared to other speech elements (e.g. diphones or demisyllables), the number of clusters is remarkably lower. Another advantage of clusters is the fact that cluster borders are usually correlated with stress borders, which reduces concatenating problems and improves the naturalnass of synthesized speech. This paper describes our transcription and phone mization techniques and some hardware aspects.