Speaker Dependent Continuous Kannada Speech Recognition Using HMM

This paper addresses the problem of Kannada speech recognition. The designed algorithm recognizes continuous Kannada speech using HMM Method and works in the speaker dependent mode. The proposed method first preprocesses the original Kannada speech signal then framing is done for every 20 millisecond with an overlapping of 6.5 millisecond. Secondly voiced part is detected through computing dynamic threshold using short time energy and magnitude of signal. Thirdly in the voiced part of signal extracts Linear-Predictive Coding (LPC) coefficients, and converts them into Real Cepstrum Coefficients. Fourthly, Real Cepstrum Coefficients are passed into k-means clustering algorithm keeping k = 3 and then passed into Baum-Welch Algorithm, using this 3 state HMM model is designed for each syllables / subwords / sentence. In this paper for experiment used 20 unique sentences which can use has commands to simple mobile sets. Each of these sentences was recorded for 10 times for training and 3 times for testing of one male speaker. The command success rate of individually uttered of sentences in experiments is excellent and has reached accuracy rate of 87.76% and miss rate of about 12.24%, the precision of 0.56, recall rate of 0.68 and F1 measure of 0.61. Computations are done using Mat lab.