Выбор UBM Модели с помощью Вариационного Байесовского Анализа для GMM-UBM Систем Распознавания Диктора

Universal background models (UBM) in the state-of-the-art speaker recognition systems are typically Gaussian mixture models (GMM). The most commonly used method for the parameter estimation of the UBM model is the maximum likelihood (ML) estimation. We propose to use the variational Bayesian analysis (VBA) instead of the ML method to estimate the parameters of the GMM. VBA helps in determining the optimal model complexity in order to avoid overfitting. Furthermore, we introduce the new criterion for fast model selection based on the values of mixture coefficients. Experiments on the male speaker data of the NIST 2006 and 2008 SRE datasets (cellular channels only) show that the speaker verification system with VBA training outperforms the system with ML training of the UBM, reducing EER approximately by 8%.

Выбор UBM Модели с помощью Вариационного Байесовского Анализа для GMM-UBM Систем Распознавания Диктора