Model-based feature compensation for robust speech recognition

This paper proposes a novel robust speech recognition approach based on the model-based feature compensation. The approach combines the GMM-based feature compensation and the HMM-based feature compensation together and employs the multiple recognition passes to achieve the best performance. In the initial recognition procedure, the GMM-based feature compensation approach is employed to give better clean model and noise model. Then we further refine these models by employing the HMM-based feature compensation approach. The statistical model of the clean speech and the noise is combined by using vector Taylor series(VTS) approximation. The experimental results show that the novel approach makes a significant improvement compared to the GMM-based feature compensation and the HMM-based feature compensation without any compensation in the initial pass.