This paper investigates the problem of robust speech recognition in additive noise. Model compensation and cepstral feature compensation techniques are evaluated and compared by experiments. The approaches considered here are MLLR and CDCN. Both approaches can be combined with CMN, which is a simple but efficient approach for robust speech recognition. Different combinations of CDCN and CMN are investigated in this paper. Noisy speech is simulated by adding different noise to clean speech with different SNR. Experiments are implemented on an isolated word recognition system. And the experimental results show that MLLR can give better performance in clean and light degraded environments, while CDCN can provide better in degraded conditions.