上下文相关的区分性模型组合的局限在于引入大的模型权重参数集,在数据有限时容易导致区分性权重训练过拟合.针对该问题,本文提出利用决策树进行上下文建模,采用最小音子错误准则构建决策树以获得最优上下文相关权重参数集.决策树构造过程中通过评估目标函数的一阶近似增量来加速最优问题集的选择,并利用精细问题集来获得更好的声学区分能力.基于多模型组合的语音识别实验表明,该方法能够增强权重训练对过拟合的鲁棒性,在大幅减小参数数量的情况下降低误识率,并优于在特征空间进行组合的方法. The limitation of the context-dependent discriminative model combination is the introduction of a large set of model weight parameters, which can lead to the over-fitting training of discriminative weights when the data is limited.In order to solve this problem, this paper proposes a decision tree for context modeling, Error criteria to construct the decision tree to obtain the optimal context-dependent weight parameter set.In the decision tree construction process, the selection of the optimal problem set is accelerated by evaluating the first-order approximate increment of the objective function and the fine set of questions is used to obtain better acoustics The experiments show that this method can enhance the robustness of weight training to over-fitting and reduce the false-positive rate when the number of parameters is significantly reduced, and is better than the combination of features in feature space Methods.