A conventional feature compensation module for robust automatic speech recognition is usually designed separately from the training of HMM parameters of the recognizer, albeit a maximum likelihood criterion might be used in both designs. In this paper, we present an environment compensated minimum classification error training approach for the joint design of the feature compensation module and the recognizer itself. By evaluating the proposed approach on Aurora2 connected digits database, a digit recognition error rate, averaged on all three test sets, of 9.15% and 13.98% is achieved for multi- and clean-condition training respectively. In comparison with the performance achieved by the baseline system without any environment compensation provided by the organizer of the ICSLP-2002 special session on Aurora tasks, our approach achieves an overall error rate reduction of 54.60%.