Fuzzy clustering based parameter clustering method in mandarin speech recognition

In a large vocabulary continuous speech recognition system, to efficiently decrease parameter size and improve the robustness of parameter training, a parameter clustering method by fuzzy clustering is proposed. Based on the structure of the phonetic decision tree, leaf nodes are used for Gaussian clustering and root nodes or shallow leaf nodes are used for covariance sharing. Experimental results show that when the number of Gaussians is reduced by 50%, recognition accuracy only decreases by 0.55%. By combining fuzzy covariance sharing, a total of 4.16% in recognition increase is achieved over the conventional system with approximately the same parameter size.