COMBINING FEATURES AT VECTOR LEVELFOR HIGHER SPEED AND ACCURACY OF SPEAKER IDENTIFICATION

Abstract

The study on combination of different feature vectors for the accuracy of Speaker Identification (SI) based on Vector Quantization (VQ) performed well when compared with other paradigms. Feature vectors based on Mel-frequency Cepstral Coefficients (MFCC) and Linear Predictive Codes (LPC) were combined. Texas Instrument and Massachusetts Institute of Technology (TIMIT) database containing 630 speakers were registered and tested for feature level combination studies. LPC feature vector manifested lower accuracy than MFCC when used as such. The combination of feature vectors through sum (MFCC+LPC) and difference (MFCC-LPC) were studied. For 42% of cases of codebook sizes studied (MFCC+LPC) gave higher accuracy than simple MFCC. The accuracy of (MFCC-LPC) combination results was better than simple MFCC for 93% of cases of codebook sizes studied. The accuracy enhancement could be used to reduce the time of speaker identification to half by using half sized codebooks of (MFCC-LPC)feature vectors with same or higher accuracy as compared to codebooks of simple MFCC feature vectors.