Corpus Supported English Text To Speech Synthesis Engine

As the international communication increases, the English language has penetrated into our daily lives. It is objectively required that the text to speech system meet the bilingual or even multilingual demands. The key to realize bilingual application is to build a synthesis engine which is adapted to non-native speakers. The synthesis engine converts the phonetic transcription into speech signals. It involves unit selection, prosody modification and speech corpus, etc. It is a kernel module in synthesis system which directly controls voice quality and prosody naturalness. Studying on the English synthesis engine, the thesis involves coarticulation of the English language, speech corpus, hybrid unit based unit selection and issues of multilingual synthesis engine. The achievements are as follows:Based on the result of speech perception experiment, the effect of coarticulation on concatenation synthesis is studied. Rules for phoneme segmentation are presented and synthesis units are identified based on the proposed concepts of "concatenate consonant" and "concatenate vowel".A multilayer and low redundancy corpus is designed in order to deal with the non-native speaker issues; the quality of synthesis engine is improved by optimizing the work of recording, segmentation and labeling; an open architecture of speech corpus is also proposed.The hybrid unit based non-uniform unit selection is proposed to utilize the coarticulation rules to enhance the synthesis quality; the prosody modification algorithm is improved; architecture of fast speech database is also presented. As the result of multilingual synthesis study, an adaptive synthesis engine is presented which is characterized by adaptation to different kinds of units, customized feature set and measure methods, shared unit selection and prosody modification algorithm, and uniform linguistic representation.