The robustness of speech recognition system in noisy environments was investigated.The distortion model in Mel-frequency cepstral coefficient(MFCC) domain is highly non-linear and difficult to deal with.A new linear distortion model was proposed by replacing the logarithm operation with its piecewise linear interpolation function.Then the esti-mation of noise parameters and compensation of acoustic models were provided.The proposed method can avoid model error introduced by utilizing linearization methods based on vector Taylor series(VTS) expansion
and significantly im-prove the robustness of recognizer in noisy environments.