Decision tree dynamic pruning method based on minimum description length in speech recognition

In decision tree based state tying, decision trees with varying tied-states denote models with varying complexity. By studying the influence of model complexity on system performance and speaker adaptation, a decision tree dynamic pruning method based on Minimum Description Length (MDL) criterion is presented. In the method, a well-trained, large-sized phonetic decision tree is selected as an initial model set, and the model complexity is computed by adding a penalty parameter which is in accordance with the amount of adaptation data. Largely attributed to the reasonable selection of the initial model and the integration of stochastic and aptotic in MDL criterion, the proposed method gains high performance by combining with speaker adaptation.