Speaker recognition with rival penalized EM training

The paper considers speaker recognition with Gaussian mixture models trained by a rival penalized EM (RPEM) algorithm. Although RPEM was applied successfully to several pattern recognition problems, our attempt to apply the algorithm in its original form to speaker recognition was not successful. We modified it by adding a discriminative threshold to prevent over penalty on mixture components, and using it with batches of feature vectors rather than the original incremental mode. We applied the modified RPEM to train speaker models with the number of Gaussian mixture components adapted individually to each speaker and used it to perform some basic speaker recognition experiments. The experiments are very reassuring about using the modified RPEM as a training method for GMM based speaker recognition. In settings with limited amount of training data, not only that the algorithm showed nice convergence to reduced order speaker models, but the resulting reduced models achieved better recognition rates than the initial higher order models.