Pronunciation Modeling for Synthesis of Low Resource Languages

Natural and intelligible Text to Speech (TTS) systems exist for a number of languages in the world today. However, there are many languages of the world, for which building TTS systems is still prohibitive, due to the lack of linguistic resources and data. Some of these languages are spoken by a large population of the world. Others are primarily spoken languages, or languages with large non-literate populations, which could benefit from speech-based systems. One of the bottlenecks in creating TTS systems in new languages is designing a frontend, which includes creating a phone set, lexicon and letter to sound rules, which contribute to the pronunciation of the system. In this thesis, we use acoustics and cross-lingual models and techniques using higher resource languages to improve the pronunciation of TTS systems in low resource languages. First, we present a grapheme-based framework that can be used to build TTS systems for most languages of the world that have a written form. Such systems either treat graphemes as phonemes or assign a single pronunciation to each grapheme, which may not be completely accurate for languages with ambiguities in their written forms. We improve the pronunciation of grapheme-based voices implicitly by using better modeling techniques. We automatically discover letter-to-sound rules such as schwa deletion using related higher resource languages. We also disambiguate homographs in lexicons in dialects of Arabic to improve the pronunciation of TTS systems. We show that phoneme-like features derived using Articulatory Features may be useful for improving grapheme-based voices. We present a preliminary framework addressing the problem of synthesizing Code Mixed text found often in Social Media. Lastly, we use acoustics and cross-lingual techniques to automatically derive written forms for building TTS systems for languages without a standardized orthography.