Speech synthesis using generative adversarial network for improving readability of Hindi words to recuperate from dyslexia

Abstract

Children learn and develop their abilities at their own pace. One of the most basic skills that they acquire is reading. However, some children struggle with reading longer than their friends, and in such a case, it is possible that they have a learning disorder known as dyslexia. The paper aims to use neural networks, namely generative neural networks, for generating raw audio data of two- or three-letter Hindi words. Using the generated data, a system will be built that will pronounce generated words for children recuperating from dyslexia. The system aims to be an effective helping tool for teachers to speed up the recuperation process by making the child repeat the correct pronunciation of the word. The system uses advance Mel-generative adversarial network neural network for working with Mel-spectrograms of the raw audio, by which the system will model its own audio iteratively, until a satisfactory result is achieved. Generated audio sample contains the Hindi words which will be taught to children. Mel-generative adversarial network will be used to generate audio samples since it provides better results compared to other existing models. 300 basic two- or three-letter Hindi words are taken as an input for assisting 5- to 8-year children. Minimum opinion score is calculated for comparison.