Abstract: In real environments the performance of speech recognition system may be significantly degraded because of the mismatch between the training and testing conditions.Model adaptation is an efficient approach that could reduce this mismatch,which adapts model parameters to new conditions by a small amount of adaptation data.Maximum likelihood linear regression (MLLR) is a popular transformation-based model adaptation algorithm.However it may degrade the performance of speech recognition system when only a few data are available.In this paper,a new model adaptation using maximum likelihood sub-band linear regression (MLSLR) is presented,which divides the full channels of Mel filter bank into several sub-bands and uses linear function to approximate the relationship between training and testing mean vectors in every sub-band.The experimental results show that the proposed algorithm overcomes the sparse data problem preferably and requires only a small amount of data.Therefore,it is more useful for rapid model adaptation.