We proposed an augmented cepstral mean normalization algorithm that differentiates noise and speech during normalization, and computes a different mean for each. The new procedure reduced the error rate slightly for the case of same environment testing, and significantly reduced the error rate by 25% when an environmental mismatch exists over the case of standard cepstral mean normalization.