Efficient Speech To Text Conversion System for Odia Language

Abstract

A human being persists many intellectual senses, among them speech is the most newlineefficient channel for communication. For the last few decades, the Speech-to-text (STT) newlineconversion system is extensively used in several application areas including industry and newlineacademia. The challenging task in speech processing is, how to recognize the speech for a newlineparticular language by a machine. Speech recognition has attracted researchers across the newlineglobe to develop systems for their native languages. In India, the ASR (Automatic Speech newlineRecognition) systems have evolved for different languages like Hindi, Bengali, Punjabi, newlineTamil, and Marathi etc. Even though considerable work can be found in many Indian newlinelanguages, the Odia language is not well explored yet by researchers, in the perspective of newlinespeech recognition. So, the unavailability of an active STT conversion system for Odia newlinelanguage and regional consequence takes refreshed for working in that language. newlineAutomatic spoken digit recognition is one of the common usages for designing a newlinevoiced dialler system. The system is typically significant for physically challenged (blind newlinepeople) or elderly people for having a telephonic conversation without substantially newlinedialling the numbers. An ample amount of literature has proposed to recognize digits for newlinedifferent languages. But less work was found for the Odia language. So, we have newlineconsidered our first work recognizing isolated voiced Odia digits using HTK (Hidden newlineMarkov Toolkit) toolkit. The proposed model for Odia digit recognition is established in newlinetwo steps: training and testing the digit model. In the training process, for each Odia digit, newlineone model is created. The training process includes preprocessing, feature extraction using newlineMFCC (Mel-frequency Cepstral Coefficient) and building the HMM (Hidden Markov newlineModel) model. In the testing phase, pre-processing and feature extraction followed by newlinerecognition has accomplished using the closest HMM model. The work is further newlineimplemented using the support vector machine (SVM) approach. Using