Factor analysis method offers state-of-the-artperformance in speaker identification during the paper. Thecompact representations of speakers named i-vectors areextracted from the utterances in a new low dimensionalspeaker- and channel-dependent space, named a totalvariability space. LBG algorithm is combined with fuzzytheory in the initialization of speaker models,whichimproves the recognition rate of the system. Channelcompensation techniques, such as Linear DiscriminateAnalysis (LDA), Principal Component Analysis (PCA),Nuisance Attribute Projection (NAP) and Within-classCovariance Normalization (WCCN) are compared duringthe experiment. It can be seen that LDA followed by WCCNachieves satisfying performance. In addition, severalidentification methods are contrasted in the experiments.One is through Support-Vector-Machine (SVM), anotherone directly uses the cosine distance similarity (CDS) as thefinal decision score, logarithmic likelihood and vectorquantization are used to compare to above two methods. Itdemonstrates that CDS combined with score normalizationobtains better result. The testing of mobile phone databaseshows the robustness of the system in complex channelenvironment. The graphical user interface of training andtesting module is simulated on MATLAB in the end of thepaper.
展开▼
