In noisy speech recognition, Wiener filters may be applied either directly to the noisy speech or, alternatively, the filters can be used to adapt the HMM mean cepstral vectors. In this paper we present experimental results which demonstrate that Wiener filters used for the adaption of HMM cepstral means perform better than the direct application of Wiener filters to the noisy signal. Experiments indicate that the variance of the cepstral features decrease with increasing noise. A theoretical explanation of the decrease in variance is presented.