Improved multitaper PNCC feature for robust speaker verification

A major challenge for practical speaker verification is the significant performance degradation in noisy circumstances. To solve the problem, a novel feature is proposed in this paper based on power normalized cepstral coefficients (PNCC). In the new feature, multitaper method and compressive gammachirp filter-bank (cGCFB) are introduced. Multitaper method provides a better spectrum estimation while cGCFB makes the sub-band power integration more robust. The experiments shows that with a suitable configuration, the improved multitaper PNCC (imPNCC) outperforms the original PNCC under both clean and noisy conditions, and achieves the best overall results on NIST 2008 and 2010 SRE database compared to PLP and PNCC.