An algorithm is proposed to build large, highly detailed acoustic models for context dependent units using a limited amount of training data. Robustness of the parameter estimates in face of data sparsity is addressed by using MAP distribution smoothing. Context dependent distributions are first clustered using a decision tree-based algorithm with an ML objective. These decision trees are then extended using a MAP objective. Experimental results show an absolute reduction in the word error rate of 0.7% by extending an existing state of the art ML trained context dependent model.