FCM BP based parameter clustering method in speech recognition

To efficiently decrease the parameter size and improve the robustness of parameter training, a parameter clustering method based on FCM BP fuzzy clustering analysis is proposed. Based on the structure of phonetic decision tree in state tying, leaf nodes are used for Gaussian clustering and root node or temporary parent nodes are used for covariance sharing. The experimental results show when the number of Gaussians is reduced by 50%, the recognition rate only decreases by 0.55%. By combining covariance sharing, a total of 4.16% recognition increasing is achieved over the conventional system with approximately the same parameter size.