Background noise can cause severe degradation of performance for speech recognition systems. Robustness towards background noise can be achieved by applying model-based compensation approaches. For systems that use MFCC features, the relationship between noise, speech, and resulting noise-corrupted speech is non-linear, and an important aspect of model-based approaches is how to approximate this relationship. To investigate how accurate s uch approximations need to be, in order to achieve good recognition performance, we apply three different techniques. These are evaluated on a spoken digit recognition task with artificially added noise.