Model adaptation is an effective approach to robust speech recognition. In this paper, we present a new approach for adaptation which is evolved from Vector Taylor Series(VTS). It improve the traditional VTS by considering the connection between the delta features and HMM parameters. Here a new method is proposed to estimate environmental parameters used in the new approach. Experimental results demonstrate that even using a small quantity of training data from the test environment, higher recognition accuracy can be achieved, compared with traditional VTS.
展开▼