Effect of MFCC normalization on vector quantization based speaker identification

Abstract

Mel Frequency Cepstral Coefficients (MFCC) are widely used in speech recognition and speaker identification. MFCC features are usually pre-processed before being used for recognition. One of these pre-processing is creating delta and delta-delta coefficients and append them to MFCC to create feature vector. Another pre-processing is coefficients mean normalization. In this paper, the effect of these two processes on the accuracy of a Vector Quantization (VQ) speaker identification system is compared. Additionally, it is shown that coefficient variance normalization, which is less common, can improve the accuracy.