语音识别中基于最小描述长度准则的决策树动态剪枝算法

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

在基于语音学决策树状态聚类时,包含不同数量捆绑状态的决策树对应不同的复杂度。通过研究模型的复杂度对系统性能和说话人自适应的影响,提出一种决策树剪枝方法——基于最小描述长度(Minimum Description Length:MDL)准则的决策树动态剪枝。该方法利用训练充分的决策树作为初始模型,根据自适应语料的数量动态地选择不同复杂度的模型,决策树剪枝时初始模型的合理选择,自适应语料的充分应用以及MDL准则对随机模型和确定性模型的集成,使得所提出的方法与说话人自适应相结合后取得了系统性能明显提高。

In decision tree based state tying, decision trees with varying tied-states denote models with varying com- plexity. 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.