Speech Mel Frequency Cepstral Coefficient feature classification using multi level support vector machine

Abstract

This paper presents the combined application of machine learning algorithm MLSVM (multi level support vector machine) and feature vector MFCC (Mel Frequency Cepstral Coefficient) to improve the result of speech recognition in comparison to other algorithm which involve formant as feature vector and KNN (k-nearest neighbour), Tree, LDA (linear discriminant analysis) and QDA (quadrature discriminant analysis) as machine learning algorithms. Problem with a machine learning algorithm like KNN, Tree, LDA and QDA is that the accuracy of recognition is limited to increase in the database and training set percentage with respect to total a database available. In this paper analysis and comparison is done between feature vector inspired by speech generation and hearing model to improve the result of recognition for larger database of twenty words. The twenty words are used to generate forty commands which are sufficient to control the on and off state of twenty homes or office appliances. The paper also presents the brief analysis of the effect of training set percentage on the accuracy.