Model Adaptation

Abstract

Both HMMs and n-gram models are normally created by using some sample set for training and are afterwards applied for the segmentation of new data. This is by definition not part of the training samples and can never be in practical applications. The characteristic properties of this test data can thus only to a limited extent be predicted on the basis of the training material. Therefore, in general differences between training and testing material will occur that can not be captured by the statistical models created and in the end adversely affect the quality of the results achieved.