TamilSpeech: Advancing TTS with Cutting-Edge Deep Learning Technologies

Abstract

"This research focuses on the text to speech system for Tamil language with cutting edge deep learning technologies. The goal is to synthesize tamil speech for the given tamil text. Text-to-Speech (TTS) systems have demonstrated their immense value by transforming written text into audible speech, providing benefits across diverse user groups. The emphasis on Tamil arises from the necessity to embrace linguistic diversity. Most significantly, Text-to-Speech (TTS) technology plays a vital role in aiding visually impaired individuals to navigate digital devices and manage their daily lives. Tamil is an ancient classical language spoken by 15% of Sri Lanka's population and the worldwide Tamil-speaking population is estimated to be around 80 million, making it one of the most widely spoken languages in the world. It poses challenges in automated speech synthesis due to its unique characters and diverse phonetic sounds. The motivation of this study is due to the lack of research on Tamil text to speech applications with cutting edge deep learning technologies. The research involves the implementation of advanced TTS models, including Parallel Tacotron 2 and Hifi GAN, to enhance the synthesis of natural-sounding Tamil speech. The models were built using the PyTorch neural network after thorough analysis of literature of the patent studies. Two models were built on the data to perform two different roles. Text encoder model was built using Parallel Tacotron 2 and speech decoder model was built using Hifi GAN model to generate speech waveform. A user interface is created using the Flask application to implement this solution. The unit testing is done for this work. This report contains the training and validation loss evaluation as a part of evaluation."