Speaker Identification Based On MFCC and IMFCC Using GMM-UBM

Abstract

In Speaker identification (SI) systems long-lasting feature extraction unit is required. For the purpose of proper representation of this features, there is a speaker modeling scheme after extraction unit. Several feature sets are used for speaker related application, one of the standard feature set is MFCC (Mel Frequency Cepstral Coefficient) which are generally modeled on human auditory system. On the other hand complementary information present in the higher frequency range known as an IMFCC (Inverted Mel Frequency Cepstral Coefficient) is another feature set which is useful. This paper concentrates on Gaussian Mixture Model (GMM)along with the Universal Background Model (UBM) is being used for the modeling purpose. Instead of triangular filters here Gaussian shaped filters are being used. Here , in this paper the results are being verified by the standard database TIMIT. The accuracy for the individual set of features such as for MFCC is coming to be 96.6% for the 16 mixtures while for the IMFCC is 95.4% for the 16 mixtures respectively for first set of speakers used and on other side the accuracy for MFCC and IMFCC in the other set of speakers is 97.22% and 86.11 respectively.