I ABSTRACT Abstract Prevailing speech recognition system could obtain a good recognition accuracy for clean speech data, but their performance will degrade rapidly in noisy environments owing to the mismatch between the acoustic models and the testing speech. There- fore, noise robust speech recognition as a hot issue receives wide attention and has been extensively studied .However, these researches can not fully meet the practical requirement due to the complexity of the noisy environments. This thesis investigates a novel and more efficient extension of GVP-HMM that can also model the trajectories of feature space linear transforms. The transforms are trained under the constrained maximum likelihood linear regression (CMLLR) criterion and would be applied on features with auxiliary information to eliminate the mismatch from clean model and test environment. In this thesis, the theories of GVP-HMM and derivation of the featur