Impact of voice excitation features on speaker verification

Mel-frequency cepstrum coefficients (MFCC) have been the most popular features used in speaker recognition. It has been recently shown that residual signal estimated through linear prediction (LP) also conveys speaker-specific information, and applied to speaker identification. In this paper, we investigate on the impact of LP-residual cepstrum coefficients (LPRC) on speaker verification along with MFCC and linear predictive cepstrum coefficients (LPCC) as well, and make comparisons of their performance in verification by conducting experiments on NIST 2001 SRE corpus, including modern classifiers. It is shown that LPRC features are as useful as MFCC and LPCC features in speaker verification, and fusing the LPRC, LPCC, and MFCC features in pairs improves the verification performance.