The Festvox Indic Frontend for Grapheme-to-Phoneme Conversion

Text-to-Speech (TTS) systems convert text into phonetic pronunciations which are then processed by Acoustic Models. TTS frontends typically include text processing, lexical lookup and Grapheme-to-Phoneme (g2p) conversion stages. This paper describes the design and implementation of the Indic frontend, which provides explicit support for many major Indian languages, along with a unified framework with easy extensibility for other Indian languages. The Indic frontend handles many phenomena common to Indian languages such as schwa deletion, contextual nasalization, and voicing. It also handles multi-script synthesis between various Indian-language scripts and English. We describe experiments comparing the quality of TTS systems built using the Indic frontend to grapheme-based systems. While this frontend was designed keeping TTS in mind, it can also be used as a general g2p system for Automatic Speech Recognition.