Noise robust speaker identification by dividing MFCC
Kizuki Matsumoto, Noboru Hayasaka, Youji Iiguni · 2014
Until now, systems using speaker identification have not been widely used. The main reason is because the identification accuracy is low. Therefore, in this paper, we report the results of modulation frequency analysis and propose a novel method using effectively modulation frequency components of vocal tract characteristics. We investigated effective components of interframe variation of MFCC. The method that we propose extracts those components and uses them effectively by dividing MFCC. The bandwidth of the division, from 0 to 7.80 Hz and from 7.80 to 25.0 Hz, was determined by the view of the modulation frequency analysis and human speech perception. We divided MFCC into two parts, low-MFCC and middle-MFCC, by the bandwidth. Additionally, we made an identification experiments on those MFCCs. Using those MFCCs improved the identification accuracy by 2.1% on average. However, the effect of the proposed method fluctuated by uttered text.