A GMM-based robust incremental adaptation with a forgetting factor for speaker verification

초록

Speaker recognition (SR) system uses a speaker model-adaptation method with testing sets to obtain a high performance. However, in the conventional adaptation method, when new data contain outliers, such as a noise or a change in utterance, an inaccurate speaker model results. As time elapses, the rate at which new data are adapted is reduced. The proposed method uses robust incremental adaptation (RIA) to reduce the effects of outliers and uses a forgetting factor to maintain the adaptive rate of new data in a Gaussian mixture model (GMM). Experimental results from a data set gathered over seven months show that the proposed algorithm is robust against outliers and maintains the adaptive rate of new data. © 2010 Springer-Verlag Berlin Heidelberg.