RECOGNITION OF SPOKEN HINDI WORDS

Abstract

Work on speech recognition is not new to our times. For many years people have been trying to make our machines hear, understand and also speak our natural language. This arduous task can be classified into three relatively smaller tasks. 1. Speech recognition to allow the machine to catch words phrases and sentences that we speak. 2. Natural language processing to allow the machine to understand what we speak, and 3. Speech synthesis to allow machines to speak. The work described in this dissertation falls under the first category. Although a complete implementation of the first part would require that the computer be able to catch words and sentences from continuous and natural speech over broad variations in accents, ages etc., to simplify the task, this work is focused on speaker dependent, isolated word recognition. The dissertation is an experimental speaker-dependent, real-time, isolated word recognizer for Hindi words. It uses Linear Predictive coding derived weighted Cepstral coefficients as the feature (observation) vectors and Hidden Marlov Modeling for recognition. Since this technique is speaker dependent therefore it can be used in applications like Voice activated Windows command controls, Home speech based telephone dialing etc.