A new score normalization method for speaker verification with virtual impostor model

User authentication system using a smart card or authentication token is drawing more and more attention. The limitation of the size of the RAM inside the processor for user authentication necessitates a verification algorithm with the efficient usage of the memory. In this paper, we present a score normalization method, which is suitable for embedded speaker verification system. Proposed score normalization method does not need physical or actual impostor model but utilizes the client model’s mean and covariance vector to imitate an impostor model. Therefore, no additional memory for impostor model is required. Furthermore, most parts of score evaluation process for impostor model are identical with that of client model. As a result, computational burden for virtual cohort model is trivial. We evaluate the performance of our proposed system in seven channel conditions and two channel compensation methods. In our experiment, the performance of our proposed speaker verification system is always superior to un-normalized likelihood based system in various channel condition. And, when comparing with un-normalized likelihood score based speaker verification system, average error rate reduction of the proposed system is higher than 6.7%.