Ideal GMM parameters & posterior log likelihood for speaker verification

Personal Identity Verification using non-alterable bio-characters is fast becoming a standard add-on layer of security for access to sensitive information. This paper presents a performance centric study of our text-independent speaker verification system using the highly parametric Gaussian Mixture Modeling [GMM] technique on the KING Speaker Verification database. The system was evaluated based on Equal Error Rate [EER] scores using different parameters - namely, the number of mixtures [M] and dimensions [D], applied to the Gaussian Mixture Model [GMM]. A new scoring method, in the quest for better performance in terms of EER is also discussed. These techniques aim to pre-determine the optimum values for M and D, and apply a scoring technique that provides optimum tradeoff between complexity and performance. Simulation results obtained confirm the implementation of speaker verification algorithms with minimal real-time adaptability requirements. This aids the development of more robust and predictive voice based authentication applications. Finally, we demonstrate that the proposed Posterior Log-Likelihood based scoring does not provide significant performance gains over log-likelihood based scoring techniques.