In this paper, we present a method for automatic generation of acoustic models from simple generic models. This method use the internal structure of non-contextual acoustic models in order to build new specialized states which are supposed to modelize specific patterns of a phoneme. The proposed technique use temporal information for state splitting. This method is compared to a maximum likelihood based approach. Our experiments show that this last criterion leads to better performance. Nevertheless, unsupervised model splitting seems to be less efficient than model specialization based on a priori knowledge.