Similarity normalization method based on world model and a posteriori probability for speaker verification

For the task of speaker verification, similarity measure normalization methods are relevant to cope with variability problems and with data and/or decision fusion issues. The aim of this paper is to suggest a new normalization method, which combines classical world model-based normalization techniques with a posteriori probability-based ones. This method presents the wellknown advantages of the a posteriori probability-based methods without requiring data and speaker specific processing. Here, it is experimented through a temporalsegmental, multi-recognizer speaker verification system. The results obtained on a subset of the SwitchboardNist98 database demonstrate the ability of this method to normalize similarity measures (in probability domain) without decreasing performance. The second advantage of this method is borne out by the performance of the multi-recognizer system, which reveals that this normalization is able to make the fusion step easier without requiring any weighting function even if individual recognizer performance is dissimilar.