Speech Quality Enhancement Using Phoneme with Cepstrum Variation Features

Abstract

In recent years, Text-to-Speech (TTS) synthesis is taking a new dimen-sion. People prefer voice embedded toys, online buyers are interested in interac-tive chat application in the form of text-to-speech facility, screen readers for visually challenged people, and many more applications use TTS module. TTSis a system that is capable of converting the arbitrary text input into natural sounding speech. It’s success lies in producing more human like speech sounding more nat-ural. The most importanttechnical aspect of TTS is feature extraction process. Both text and speech features are needed but it is not that easy to select meaning-ful and useful features from the text or from speech. There are many feature extraction techniques available for both text and speech, still there is a need for very simplest form of feature extraction technique. Though the emergence of Deep learning technique automates feature extraction, it is suitable only when the volume of data is enormous. This paper proposes a novel text and speech feature extraction technique which is based on special symbols present in the text and phoneme with cepstrum variation of the speech signal respectively. These techniques are simple and works well for real-time applications in which size of data is small or moderate. The proposed methods not only extract useful features but also mean-ingful features in terms of fetching the salient traits of the text and speech cepstrum. The experimental results have shown that the quality of speech is increased by 14% when compared to the other conevntional feature extraction techniques.