Gaussian mixture model - support vector machine (GMM-SVM) with nuisance attribute projection (NAP) has been found to be effective and reliable for speaker and language recognition. In maximum a posteriori (MAP) adaptation of GMM, the relevance factor is the parameter that regulates how much the adaptation data affect the base model, which impacts the final recognition performance. In our previous work, the data-dependent relevance factor and adaptive relevance factor have been introduced. In this paper, we provide insights into different types of relevance factor for MAP in the context of application as formulated under Speaker Recognition Evaluation (SRE) and Language Recognition Evaluation (LRE) by the National Institute of Standards and Technology (NIST).
Index Terms: maximum a posteriori, supervector, Gaussian mixture model, support vector machine