Comparison of vector normalization methods in multi-level speaker verification

In this article a text-independent speaker verification problem is considered. After the feature extraction, each conversation side has been represented as a vector in a fixed dimensional space. In order to reduce an influence of the lengths of utterances and also the channel properties, various vector normalization techniques have been selected from the literature, modified, and tested. Additionally, it is shown that if the vectors are transformed in such a way that they lie on the unit hypersphere, better numerical properties of the respective kernel matrices can be achieved. Finally, various normalization methods as well for continuous features (i.e., the cosine kernel, the variance normalization, the cosine similarity merged with the variance normalization, and the variance normalization merged with the spherical normalization) as for the discrete features (again the cosine kernel, the TFLOG, the TFLOG merged with the cosine similarity metric, and the TFLOG merged with the spherical normalization) are tested and compared in this article.