You will have to consider the sources of sound, and each source has its own distribution. The resulting sound is a mixture of all the sound sources.
In the case of the car on a raceway track, we might consider the following sound sources: Wind, tyres, engine, human (in the car).
Wind and tyres I would expect to have sound level correlated strongly with the speed of the car. I would expect them to be almost white noise, regardless of speed. However the engine sound I would expect to also be correlated with the acceleration. I expect that the engine sound, and as engine RPM goes up, frequency content shifts upwards too. Assuming that the car has gears, engine RPM is non-linear with speed. Human is maybe uncorrelated to speed,acceleration.
All these expectations must be checked according to the data.
To eliminate the uncorrelated effects, you will need many samples at multiple points of speed,accelerations,RPMs. To simplify the problem you can try to keep acceleration and RPM constant.
If you are OK with simple metrics like overall sound level and spectral mean, you can use ordinary curve fitting on the observations. Note that the correlation might not be linear! You will more than 3 measurement points for this to make sense.
Once that simple model is in place, you can try to model the MFCCs more directly. Try a regression model that supports non-linear relationships, like RandomForest or kernel SVM. Also consider a dimensionality reduction like PCA, as MFCC values are likely to be correlated with eachother. As long as you avoid overfitting, you can use such a model to interpolate between speeds.