Abstract: A new rapid speaker adaptation method based on maximum likelihood variable subspace is proposed. A set of bases of the speaker space is obtained by performing Principal Component Analysis (PCA) on the Speaker Dependent (SD) model parameters of the training speakers. Different from conventional subspace based methods, during speaker adaptation, a subset of these bases is dynamically chosen for each speaker using maximum likelihood criteria. The new speakers model is constrained in the subspace spanned by those bases. With less free parameters required, the new method can obtain more robust SD model using very little amount of adaptation data. Speech recognition experiments show that the new method can obtain better performance than the eigenvoice method and MLLR method, both in supervised mode and in unsupervised mode.