Implementation and performance evaluation of continuous Hindi speech recognition

Speech to Text recognition is the ability of a machine to recognize the human speech and convert in to text sequence. In this paper, we compare the performance of isolated word, connected word, and continuous speech recognition system with different vocabulary sizes. Hidden Markov Model toolkit HTK 3.4.1 is used to develop the system. For feature extraction, Mel Frequency Cepstral Coefficient (MFCC) and Perceptual Linear Prediction (PLP) both are used in this paper. The aim of this paper is to build a high performance speech recognition system for Hindi language. Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM) are used at the back-end of our proposed system. The system is trained for 100 Hindi words and each word 10 utterances have been recorded for training of the ASR system. The experimental result shows that the overall accuracy of proposed system with 100 word dictionary size is 95.40%, when we use the combination of MFCC and GMM for automatic speech recognition (ASR) system.