Several techniques for the adaptation of DTW and HMM speech recognizers to car environment are presented and tested. These techniques are : nonlinear spectral subtraction, spectral subtraction with neural networks, two strategies of feature parameters transformation using linear regression or neural networks, and finaly two approaches to adapt directly the HMM states distributions. Results given prove the superiority of parameters transformations on spectral subtraction, neural networks transformation on linear regression transformations, and adding noise to references on speech enhancement strategy. They prove also the importance of the choice of the spectral representation and the corresponding distance measure. A number of perspectives exists : defining adaptative transformations, transformations by class of noise, adapting the HMM parameters using NN.