Gaussian mixture model compensation method using shift factor for speaker recognition

The performance of GMM-based text-independent speaker recognition systems declines rapidly when the training data is reduced. A model compensation method is proposed to address the problem. Since there is a shift between each target GMM-based model and the UBM (Universal Background Model), a low-dimensional affine space is fined, named shift space, and the shift for each model with sufficient training data is transformed to the shift factor in this space. When the training data of the target speaker is insufficient, firstly, the coordinate of the shift factor is learned from the GMM mixtures of insensitive to the amount of training data, and then it is adopted to compensate other GMM mixtures. Using the proposed method, a relative reduction of 7% in EER (equal error rate) is obtained comparing with the baseline system.