Synthetic vehicle sounds can be used for subjective listening tests. Computer models provide a great deal of flexibility while mechanical systems are not easily modified for such tests. Autoregressive (AR) parametric models were developed using real data for training. With this approach relatively few parameters are needed to model the vehicle sounds. Furthermore, arbitrarily long data sequences can be produced from these models for subjective evaluations by listeners. Interior recordings in two 4‐cylinder engine cars and in two 6‐cylinder engine cars were used as training data. It was found that signals with more complex spectra were modeled most successfully from the point of view of a listener. Signals with simple spectra (one dominant frequency component) were more difficult to model. Modeling difficulty manifests itself in amplitude variability with time in the sounds produced with the synthetic data. Measures to counter these effects have been met with some success. [Work supported by Ford Motor Company.]