This paper deals with the development of an isolated-word speech recognition system operating in the car environment. The problems related to the limited size of training data and to the different operating conditions between training and testing are addressed, and the performance of the system reported.