Optimal splitting of HMM Gaussian mixture components with MMIE training

摘要: A novel approach to splitting Gaussian mixture components based on the use of maximum mutual information estimation (MMIE) training is proposed. The idea increase acoustic resolution only in those distributions where discrimination problems are identified. Problem determined by looking at each weight counter; a large positive counter value indicates both that component often tends not be recognized correctly (i.e., part best path when it should be) and there sufficient data split component. Results a, connected digit recognition experiment TIDIGITS corpus indicate much better results can obtained with such MMIE trained models than MLE several times more components.