This paper carried out a subjective evaluation test with magnitude estimation for 78 noise samples to evaluate the sound quality of vehicles. In the test, six types of B-Class vehicles were taken as the study objects and sound signals collected in co-driver locations at steady states as experimental samples. Meanwhile, seven objective parameters were calculated to describe the sound characteristics. By using objective parameters as inputs, subjective values as outputs, a GA-BP neural network was adopted to establish a sound quality prediction model. Experiments show that the model gives good predictions of high correlation (0.928) and low error (±8%). Then, the network connection coefficients were used to calculate the impact weight of objective parameters on the results of subjective evaluation, and a new model with main parameters was established. As expected, the loudness, sharpness and roughness with a total relative importance of 83% are the most influential parameters in vehicle interior sound quality.