Letter-to-Phoneme Conversion for a German Text-to-Speech System
Vera Demberg · 2006
This thesis deals with the conversion from letters to phonemes, syllabification and word stress assignment for a German text-to-speech system. In the first part of the thesis (chapter 5), several alternative approaches for morphological segmentation are analysed and the benefit of such a morphological preprocessing component is evaluated with respect to the grapheme-to-phoneme conversion algorithm used. In the second part (chapters 6- 8), a pipeline architecture in which syllabification, stress assignment and letter-to-phoneme conversion are tackled in separate modules is compared with a design that integrates all tasks into one step. Alternative approaches to the problems are discussed and evaluated. Best results are obtained with a Hidden Markov Model that does letter-to-phoneme conversion, syllabification and stress assignment simultaneously and incorporates some very simple and language-independent phonological constraints regarding syllable structure and word stress. This model achieves a 50 % reduction in word error rate relative to the AWT decision tree. The Hidden Markov Model is shown to compare well to state-of-the-art approaches to grapheme-to-phoneme conversion in English and French.