Marathi connected word speech recognition system

The manuscript presents Marathi connected word speech recognition system. The system is implemented using Mel-frequency Cepstral Coefficient (MFCC) feature extraction technique and Continuous Density Hidden Markov Model (CDHMM). Initially, recorded speech signals of connected words are segmented into words using speech segmentation algorithm. The speech segmentation algorithm is based on two basic features of speech signal i.e. Short Time Energy (STE) and Spectral Centroide (SC). The features of each isolated words are extracted with MFCC. The extracted features are used to develop CDHMM for each word present in the speech signal. The multivariate Gaussian mixture density function is used to define the observation probabilities generated by each state. The Baum-Welch algorithm is used to re-estimate the model parameters. The Bi-gram language models are used to find Bi-gram pairs. Finally, the log Viterbi beam search is used to find the maximum likelihood state sequence path. The experimental result shows the overall system accuracy.