Sub-band based text-dependent speaker verification

, (p s c S p th cepstral coefficient of the s th sub-bands { c 1 (1,p) = c(p) is the p th full-band cepstral parameter} S number of sub-bands Y(k) k th log spectral magnitude K number of log spectral magnitudes) (k Y ′ ′ k th log-energy outputs of the mel-scale filterbank K ′ ′ number of log-energy outputs of the mel-scale filterbank h t weight associated with the t th segment U number of competing speakers (1−η s) average speaker verification error rate in the s th sub-band R, q number of sub-band-systems Superscript s implies the association of the s th sub-band Superscript u implies the association of the u th competing speaker Superscript r implies the association of the r th sub-band-system ′ implies the association of the complementary full-band cepstral parameters SB-Sub-band based DBSW dynamic band-limited segmental weights MFBO mel-scale filterbank output SNRW signal-to-noise ratio-based weighting factors CFBCC complementary full-band cepstral coefficient SSFCS sub-band system with full-band cepstral supplements MSBSA multiple sub-band-systems analysis FWN filtered white noise RNT real noise type Abstract This paper addresses various issues involved in sub-band based text-dependent speaker verification. The first part of the discussions is concerned with the classification methods. An important issue addressed in this part is the determination of a set of weights which emphasises the sub-bands that are specific to the target speaker while de-emphasising or removing the contaminated ones. In particular, techniques for determining these weights dynamically according to the level of contamination in the sub-bands are described. Furthermore, the effectiveness of these methods is experimentally analysed through a set of comparative studies. The second part of the discussions focuses on the feature extraction process. Analytically, it is shown that for a sub-band system of S bands, the cepstral coefficients with the quefrency of p have a strong linear relationship to the (S×p) th full-band cepstral parameter. With the aid of a set of experimental results, it is demonstrated that this means the conventional classification methods adapted to work with sub-band cepstral parameters may not be able to capture all the useful spectral information contained in the full-band cepstral parameters. In order to tackle this problem, two methods are described and their relative effectiveness is experimentally examined. The experimental investigations also include an examination of speaker discrimination abilities of different sub-bands and an analysis of different possible recombination levels.