Recently, we, at The University of Hong Kong (HKU) have proposed several approaches based on stochastic vector mapping and switching linear Gaussian HMMs to compensate for environmental distortions in robust speech recognition. In this paper, we present a comparative study of these algorithms and report results of performance evaluation on Aurora connected digits databases. By following the protocol specified by the organizer of the Eurospeech-2003 special session on Aurora tasks, the best performance we achieved on Aurora2 database is a digit recognition error rate, averaged on all three test sets, of 5.53% and 6.28% for multi- and clean-condition training respectively. In a preliminary evaluation on Aurora3 Finnish and Spanish databases, significant performance improvement is also achieved by our approach.