The implementation of PFCMS using cepstrum information

In this paper, we propose a method to remove the residual speech effects of the cepstral mean used in channel normalization. Rather than using the roots of the linear prediction polynomial, we propose the pole-filtering method that utilizes the pre-determined cepstral information for the implementation of Pole-Filtered Cepstral Mean Subtraction (PFCMS), a method that effectively separates speech information from the cepstral mean of channeled speech. Using 8 simulated telephone channels, we compared the relative errors of estimating channel cepstrum, speaker identification and computational efficiency for CMS, PFCMS and the proposed method respectively on two databases. The proposed method, the Formant-Broadened CMS (FBCMS) has shown to yield improved speaker recognition rates with lower computational burden.