AbstractsLanguage, Literature & Linguistics

This thesis focuses on the phonetic transcription in the framework of text-to-speech conversion, especially on improving adaptability, reliability and multilingual support in the phonetic module. The language is constantly evolving making the adaptability one of major concerns in phonetic transcription. The phonetic transcription has been addressed from a data- based approach. On one hand, several classifiers such as Decision Trees, Finite State Transducers, Hidden Markov Models were studied and applied to the grapheme-to-phoneme conversion task. In addition, we analyzed a method of generation of pronunciation by analogy, considering different strategies. Further improvements were obtained by means of application of the transformation-based error-driven learning algorithm. The most significant improvements were obtained for classifiers with higher error rates. The experimental results show that the adaptability of phonetic module was improved, having obtained word error rates as low as 12% (for English). Next, steps were taken towards increasing reliability of the output of the phonetic module. Although, the G2P results were quite good, in order to achieve a higher level of reliability we propose using dictionary fusion. The ways the pronunciations are represented in different lexica depend on many factors such as: expert¿s opinion, local accent specifications, phonetic alphabet chosen, assimilation level (for proper names), etc. There are often discrepancies between pronunciations of the same word found in different lexica. The fusion system is a system that learns phoneme-to-phoneme transformations and converts pronunciations from the source lexicon into pronunciations from the target lexicon. Another important part of this thesis consisted in acing the challenge of multilingualism, a phenomenon that is becoming a usual part of our daily lives. Our goal was to obtain such pronunciations for foreign inclusions that would not be totally unfamiliar either to a native or proficient speakers of the language to be adapted, or to speakers of this language with average to low proficiency. Nativization by analogy was applied to both orthographic and phonetic forms of the word. The results obtained show that phonetic analogy gives better performance than analogy in the orthographic domain for both proper names and common nouns. Both objective and perceptual results obtained show the validity of this proposal.; Fa tan sols uns deu anys les aplicacions de sistemes TTS eren molt més limitades, encara que un passat tan recent sembla més llunyà a causa dels canvis produïts en les nostres vides per la invasió massiva de les tecnologies intel·ligents. Els processos d’automatització de serveis també han assolit nous nivells. Què és el que defineix un bon sistema TTS avui dia? El mercat exigeix que aquest sigui molt adaptable a qualsevol tipus d’àmbit. També és imprescindible un alt nivell de fiabilitat ja que un simple error d’un TTS pot causar problemes seriosos en el nostre dia a dia. La nostra agenda és cada…