Objective wavelet packet features for speaker verification

Studying ways for achieving a better demarcation of human voices for the task of speaker verification and taking advantage of the flexibility provided by wavelet packet analysis, we investigate in an objective way the relative importance of constituent disjoint frequency subbands of speech signals. Based on experimental results measuring the actual contribution of these subbands in relation to the corresponding frequency resolution, we propose a novel wavelet packet-based speech feature set that is effectively designed for speaker verification. The practical significance of our approach has been evaluated in comparative experiments performed on 2001 NIST Speaker Recognition Evaluation database. The proposed wavelet packet feature set has proven to outperform the widely used Mel-frequency scaled cepstral coefficients (MFCCs), as well as other wavelet packet based features that have been successfully used for speaker recognition.