Training a speaker verification system on unlabelled data

In the article we consider a method of labeling speaker data using clusterization techniques. Labelling problems arise when one needs to use a speaker database from new channels, for example, mobile devices. Newly labelled database might then be used to construct a speaker verification system. In the article described a speaker verification task along with some methods to solve it which are based on GMM-UBM, also some channel normalization techniques are described, which might enhance the quality of recognition. Methods based on supervectors and PLDA are also presented. We also study the quality of labeling obtained through clusterization with different metrics. Resulting labelled database is then used to train several PLDA models. Then these models fused and used to solve a speaker verification task on i-vectors from NIST are i-vector Machine Learning Challenge 2014.