Speech Synthesis: An Empirical Analysis of Various Techniques in Text to Speech Generation

Speech is the most important resource of communication and a part of human development. Speech Synthesis gives an opportunity to communicate with deaf and dumb people. The state of the art in speech synthesis is automatically generate speech from text. In the digital era, a text-to-speech synthesizer system is able to read any text clearly. This is a primary difference between any other talking machine and a text-to-speech synthesizer. The recent task of speech synthesis is to produce speech in the form of acoustic waveforms from text input. It is used in conversational and nonconversational applications also. In conversational applications like speech recognizers, telephone-based dialogues with people, current chatbots, etc., and in non-conversational applications like video games, children's toys, devices, or machines which are helpful for blind people to read out the text or messages loud. The purpose of this article is to analyze generation-wise growth in speech synthesis techniques. In the comparative analysis, we have done the analysis of speech synthesis methods and a few evaluation parameters such as corpus size, the complexity of each technique, intelligibility, the naturalness of sound, advantages, and limitations. We have highlighted a few research gaps in those techniques. As a contribution proposed a framework to produce speech from the text as well as images using the Lj speech dataset. We have adopted the Tacotron model to generate speech from the text in the English language. For an evaluation generated speech is evaluated with a Mean Opinion Scale parameter from 30 listeners.