Text to speech synthesis using deep learning

The audio deep fake is a process to generate speech similar to some specific people using various methods from text utterances of natural language. The task of speech synthesis is challenging due to the unavailability of any general model that can generate speech from all existing languages. Therefore, in this chapter, we analyze the existing techniques for Text-To-Speech (TTS) synthesis considering their architecture. Moreover, we propose a novel TTS synthesizer based on a deep learning model i.e. encoder, decoder, synthesizer, combined with a reward block based on Reinforcement Learning (RL). To the best of our knowledge, our proposed synthesizer is the first TTS synthesizer that can generate speeches in three natural languages such as English, Chinese, and French. Furthermore, we performed an experiment based on Mean Opinion Scale (MOS) which showed that our proposed model is an efficient technique for TTS synthesis.