This paper presents a novel compensation approach, which is implemented in both model and feature spaces, for nonstationary noise Due to the nature of non-stationary noise which can be decomposed into constant part and residual noise part, our proposed scheme is performed in two steps: before recognition, an extended Jacobian adaptation (JA) is applied to adapt the speech models for the constant part of noise; during recognition, the power spectra of noisy speech are compensated to eliminate

more »

... effect of residual noise part of noise. As verified by the experiments performed under different stationary and non-stationary noise environments, the proposed JA is superior to the basic JA and the joint approach is better than the compensation in single space.