[Conference] Global Conference on Cyberspace (GCCS), New Delhi, India, November 2017
Arun Baby, Anju Leela Thomas, Jeena Prakash, Anusha Prakash and Hema A Murthy
In this paper we discuss a consortium efforts on building text to speech synthesis systems (TTS) for Indian languages. There are two tasks that are crucial for building TTS systems, namely, parsing, and labeling. Although Indian languages are more or less phonetic, parsing especially the issue of schwa deletion must be addressed carefully. Accurate labeling of speech at the subword level is another important task. Owing to the nonavailability of large vocabulary continuous speech recognition systems in Indian languages, accurate labeling at the subword level is a difficult task. A universal parser across all Indian languages was first developed. A novel approach to obtain accurate labels is also proposed, where signal processing cues are used in tandem with machine learning.