A speech recognition adaptation method for a vehicle having a telematics unit with an embedded speech recognition system. Speech is received and pre-processed to generate acoustic feature vectors, and an adaptation parameter is applied to the acoustic feature vectors to yield transformed acoustic feature vectors. The transformed acoustic feature vectors are decoded and a hypothesis of the speech is selected, and the adaptation parameter is trained using acoustic feature vectors from the hypothesis. The method also includes one or more of the following steps: the speech is observed for a certain characteristic and the trained adaptation parameter is saved in accordance with the certain characteristic for use in transforming feature vectors of subsequent speech having the certain characteristic; use of the trained adaptation parameter persists from one vehicle ignition cycle to the next; and use of the trained adaptation parameter is ceased upon detection of a system fault.