Speaker Verification Reinforced by Objective Wavelet Packets-based Speech Parameterization

In attempt to enhance the discrimination ability of present speaker verification systems, we study alternative ways for parameterization of human voice. Utilizing the flexibility provided by wavelet packet analysis, and rendering an account of the most recent psycho-acoustical studies, the authors investigate in an objective way the relative importance of constituent disjoint frequency subbands of speech signal. In extensive experimentation, the contribution of each subband in relation to its corresponding frequency resolution is quantified. Taking an advantage of the insight gained in these experiments, we construct a novel wavelet packet-based speech feature set that renders an account of the relative importance of the frequency subbands. In comparative experiments performed on 2001 NIST Speaker Recognition Evaluation database, contrastive evaluation of the proposed speech features, Mel-frequency cepstral coefficients (MFCCs), as well as previous wavelet packet based features that were reported successful for speaker recognition is performed.