Feature Data Reduction of MFCC Using PCA and SVD in Speech Recognition System

The development of the pattern recognition system has increased rapidly in this century. Many developments of methods have been done. Mel Frequency Cepstral Coefficients (MFCC) is a popular feature extraction method but still has many disadvantages, especially regarding the level of accuracy and the high dimensional feature of the extraction method. This paper presents the feature data reduction of MFCC using Principal Component Analysis (PCA) and Singular Value Decomposition (SVD). Combining MFCC and data reduction methods, it is expected to improve the accuracy and increase the computational speed of the classification process by decreasing the dimensions of feature data. The result of extraction MFCC feature data plus the delta coefficient forms the matrix data which will be combined with the data reduction method. The data reduction process is designed into two versions. Then the results of data reduction are done classification process with Support Vector Machine (SVM) method. The dataset is composed of 140 recorded speech data from 28 speakers. The results showed that MFCC + PCA version 2 and MFCC + SVD version 1 were able to provide the maximum accuracy improvement with an increase of accuracy from conventional MFCC method from 83.57% to 90.71%. In addition, MFCC + PCA version 2 and MFCC + SVD version 1 method can accelerate the process of classification in speech recognition system from 7.819 seconds into for about 7.6 seconds by decreasing dimension of feature data from 26 into 10 for MFCC + PCA version 2 and decreasing dimension of feature data from 26 into 14 for MFCC + SVD version 1.