The Description of the Intra-State Feature Space in Speech Recognition

In speech recognition, the description of the intra-state feature space is an important issue in systems based on HMM-derived acoustic models. The existing techniques include the famous methods based on VQ technique and mixture Gaussian densities. In this paper, a method based on sub-space division is proposed. Experiments are done to find how many densities should be used to better describe the intra-state feature space, and the experimental results show that the number of densities should depend on the particular distribution of that space and can be judged by a kind of criterion.