Abstract
This paper reports our recent efforts in the attempt to apply the relevance vector machine (RVM) to text-independent speaker recognition tasks. The RVM represents a Bayesian extension of the widely applied support vector machine (SVM), one of the leading approaches to pattern recognition and machine learning. Both the SVM and the RVM use a linear combination of kernel functions centered on a subset of the training data to make regressions or classifications. In the SVM, the number of vectors in the subset grows linearly with the size of the available training data, while in the RVM, only the most relevant vectors will be captured. So the RVM yields a much sparser approximation of the Bayesian kernel than the SVM. Our preliminary experimental results show that the RVM overall outperforms the SVM on speaker recognition while being advantageous over the latter for its exceptionally sparse nature, classification accuracy, and Bayesian probabilistic framework. Comparisons are also made for the Gaussian mixture model (GMM), a widely used non-discriminative approach to speaker recognition.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have