Text-To-Speech for Languages without an Orthography
Sukhada Palkar, Alan W. Black, Alok Parlikar · 2012
Speech synthesis models are typically built from a corpus of speech that has accurate transcriptions. However, many of the languages of the world do not have a standardized writing system. This paper is an initial attempt at building synthetic voices for such languages. It may seem useless to develop a text-to-speech system when there is no text available. But we will discuss some well defined use cases where we need these models. We will present our method to build synthetic voices from only speech data. We will present experimental results and oracle studies that show that we can automatically devise an artificial writing system for these languages, and build synthetic voices that are understandable and usable.