Kannada word recognition system using HTK

In the present work, speech recognition system for Kannada language has been implemented using the Hidden Markov Tool Kit (HTK). The system performance is comparatively studied and evaluated for syllable and phone level models. The Kannada word dictionary of size about 110 words is used in the study and Mel frequency cepstral coefficients (MFCC) are computed in acoustic front-end processing. The system is designed to recognize isolated utterances of Kannada words, which are recorded from a Kannada short story. Baum-Welch algorithm is used to train the Hidden Markov Model (HMM) and Viterbi algorithm for decoding process. The objective of this study is to compare the performances of phone-level and syllable-level acoustical models for small to medium sized Kannada language vocabulary. The results are part of the on-going research work on large vocabulary continuous speech recognition system for Kannada language. Average word recognition accuracy of 97.1% for syllable-level modeling and 98.6% for phone-level modeling has been reported. Analysis of system performance also carried out based on the confusion matrices.