I-vector based text-independent speaker identification

Factor analysis is mainly by extracting the compact representations of speakers´ utterances, which are referred to as i-vectors. A low new space called total variability space, which is speaker and channel dependent is trained in the modeling. During the experiments, channel compensation approaches are used to remove the interference included by i-vectors. They are respectively are Nuisance Attribute Projection, Linear Discriminate Analysis, and Within-Class Covariance Normalization. Results have shown that the combination of Linear Discriminate Analysis and Within-Class Covariance Normalization obtains better performance. In addition, the system contrasts two methods to estimate the similarity between the testing speaker and the target speaker. One is through Support-Vector-Machine (SVM), the other one directly uses the cosine distance similarity (CDS) as the final decision score. The results demonstrate that CDS achieves better performance. Finally, score normalization technique is used to reduce the difference caused by channel variability. The paper proved that the combination of the above methods exactly improves the robustness of the system on the basis of guaranteeing the recognition rate. The design of the identification system is simulated on MATLAB, which includes both training and testing.