汉语语音识别声调模型集成中基于决策树的上下文相关权重参数聚类方法

摘要

声调集成是汉语语音识别的一个重要任务.在语音识别的二次解码过程中,使用区分性训练的权重因子进行声调模型集成已被证明是一个有效的方法,而且使用上下文相关的得分加权进行模型组合也得到了应用.上下文相关模型组合方法的一个不足是将会带来大量的训练参数,从而导致权重训练受到过拟合的影响.针对该问题,提出利用声学决策树对上下文相关权重参数进行参数聚类,决策树节点问题集根据最小化训练数据的期望 误识率进行选择.提出问题集剪枝来加快决策树的构建速度.汉语连续语音识别实验表明与人工选择上下文相关权重参数相比,该方法能够在大大减少参数数量的条件下明显降低误识率.%Tone model integration is an important task in Mandarin speech recognition. Discriminative model weight training is an effective technique for this purpose. In recent works, context-dependent scaling is often applied for better interpolation between the models. One limitation of this approach is a large number of parameters will be introduced, which makes it prone to overtraining. In this paper, we propose parameter lying to cluster context-dependent model weights using phonetic decision trees. Question at each tree node is chosen to minimize expected error of the training data. Question set pruning is used in node splitting to make tree building efficient. Experimental results on continuous speech recognition task show the method is capable of achieve better accuracy using many fewer parameters.