I-vector Normalized Method Based Probabilistic Linear Discrimination Analysis For Speaker Verification Research

The task of a speaker verificat ion system is to determine whether the unknown voice is fro m the claimed speaker or not.Vo ice served as a bio metric authent ication technology,it has derived extensive attent ions and been applied broadly as a critical assisted technology in voice applicat ions.Talking about speaker verificat ion,i-vector based PLDA speaker verification system has attracted a wide spread attention in related field because it is benefit by its excellent verificat ion accuracy and real-time performance.For PLDA system,it has a Gaussian assumpt ion towards i-vectors distribut ion in the modeling phase;however,the assumption is so mewhat harsh when facing channel distortion and short utterance verificat ion.For verification system,a critical factor of feature modeling is the abilit y o f distribut ion simulat ion and the disabilit y o f fitt ing the distribut ion o f i-vectors in PLDA system can affect the verificat ion accuracy and stabilit y.Hence,how to so lve the contradict ion in this system is one of the crit ical problems need to reso lved.For the flexibilit y and universalit y o f the system,a non-parametric method is proposed to Gaussianizat ion i-vectors.To meet the Gaussian assumpt ion in our proposed method,kullback-leibler divergence(KLD)is adapted as a distance measurement metric and a non-linear transformat ion funct ion is proposed to reduce the related KLD.To verify the effectiveness of our proposed method,experiments are performed on the core set o f NIST 2008.Fro m the experiment al results we can conclude that the value o f KLD is related to the verificatio n accuracy and the verificat ion accuracy can be improved by reduce the value o f KLD.To solve the disadvantage of KL-DN at the condit ion o f short utterance speaker verificat ion,an iterative Gaussianization method has been proposed which st ill use KLD and non-parametric the method.In the every iterat ive of this method,a marginal Gaussianizat ion techno logy and matrix rotat ion are applied to reduce the distance of i-vectors distribut ion and Gaussian.Fro m the experiments,we can conclude that the proposed method can improve the a ccuracy of short utterances verification.