Based on confusions between hidden Markov model (HMM) states, a state-restructuring method was proposed. In the method, HMM states were restructured by sharing Gaussian components with their related states, and the re-estimation to the increased-parameters, i.e., the inter-state weights, was derived under the expectation maximization (EM) framework. Experiments were performed on speaker-independent, large vocabulary, continuous Mandarin speech recognition. Experimental results showed that the state-restructured systems outperformed the baseline, and achieve significant improvement on recognition accuracy compared with the conventional parameter-increasing method. Such comparative results confirmed that the state-restructuring method was efficient.
