Speaker Independent Isolated Kannada Word Recognizer

Abstract

This paper addresses the problem of recognizing spoken Kannada words. The designed algorithm recognizes spoken Kannada words independent of speakers. The proposed method normalizes the original speech signal of every isolated word and extracts Linear-Predictive coding (LPC) coefficients, and converts them into Real Cepstrum Coefficient. These Real Cepstrum Coefficient values are subjected to dimensionality reduction through normal fit. These coefficients are used as the representatives of each spoken word. Euclidian distance measure is then used to compute the distance between the test samples to the model data in the database. The model datum in the database at a minimum distance is declared as the recognized word. For experimentation, we have used 294 unique Kannada words. Each of these words was recorded with 10 Speakers yielding 2,940 samples in total. Out of 10 speakers’ data, 8 speakers’ data i.e., 2,352 samples were used to compute the representative co-efficient for each word. Remaining 2 speakers’ data along with re-recorded data of two speakers out of the 8 speakers is used for testing. Totally 2,352 signals are used for training and 1,176 signals are used for testing. The success rate of the proposed system- known speaker data is 98.29 % and unknown speaker data is 91.66 %.