Differential MFCC and Vector Quantization used for Real-Time Speaker Recognition System

This paper makes some improvements on MFCC feature extraction and proposes a quick MFCC algorithm which is used for Real-Time Speaker Recognition System. Based on the quick MFCC algorithm, the paper uses Differential MFCC for feature extraction and Vector Quantization plus GMM model for classification to achieve a better result. It can meet the requirements of real-time system in case of the high precision. By comparing with the traditional MFCC algorithm, the quick MFCC algorithm reduces the run time greatly while maintaining recognition accuracy of the system. To prove it, this paper compares the quick MFCC algorithm with LPC and FFT. The experiment indicates that the EER of LPC is 14.4% and the EER of FFT is 12.5% but by using the Quick MFCC the EER is 9.4% and the differential MFCC is only 6.9%.