Speech Recognition in Kannada using HTK and Julius: A Comparative Study

This paper describes a comparative study of continuous live speech recognition for selected commonly used words and phrase in the Kannada language using the HTK Toolkit, Julia Script, and the Julius live speech recognition engine. The feature vectors used in this study are the Mel Frequency Cepstral Coefficients (MFCC). A speaker independent continuous speech recognition system has been built for a vocabulary size of 115 phrases using 16 speakers for training using the Hidden Markov Modeling(HMM) technique, Baum Welch re-estimation and the Viterbi algorithm for decoding. On comparing the HTK and Julius Live recognition engine accuracies ,it is found that the Julius engine performs better under moderate noise conditions, which is explained in this work.