Coupling a generative model with a discriminative learning framework for speaker verification

The task of speaker verification (SV) is to decide whether an utterance spoken by a target or imposter speaker. In most studies SV, log-likelihood ratio (LLR) score estimated based on generative probability model features, and compared with threshold for making decision. However, the usually focuses individual feature distributions, does not have discriminative selection ability, easy be distracted nuisance features. as hypothesis test, could formulated binary discrimination where neural network learning applied. learning, features removed help label supervision. pays more attention classification boundaries, prone overfitting training set which may result in bad generalization test set. this paper, we propose hybrid framework, i.e., coupling joint Bayesian (JB) structure parameters framework SV. two-branch Siamese built dense layers that are coupled factorized affine transforms used JB model. LLR estimation according distance metric framework. By initializing generatively learned model, further train pairwise samples task. Moreover, direct evaluation (DEM) SV minimum empirical Bayes risk (EBR) designed integrated objective function learning. We carried out experiments Speakers wild (SITW) Voxceleb. Experimental results showed our proposed improved performance large margin state art models