决策树结构对说话人自适应影响的研究

Research on the influence of decision tree structure on speaker adaptation

提出一种利用自适应语料和训练语料对模型状态结构调整的算法。该算法在易混淆的状态间参数共享,提高了模型对样本的后验概率和对自适应语料的利用率,并间接地调整了系统决策树的结构。识别实验结果表明,在不同数量的自适应语句下,调整后的系统识别率比基线系统获得了一致的提高,结合使用MLLR说话人自适应,调整的系统识别率平均提高了 15．60％,有效地减少了测试语料与训练语料决策树结构不匹配造成的系统识别率降低。

Based on adaptation data and training data, a state restructuring method was proposed. In the method, parameters between confused states were shared, which improved the posterior probability, made better use of adaptation data and indirectly restructured the decision tree of the baseline. Experimental results showed when a varying number of adaptation sentences were taken from each speaker, restructured system increased recognition rate consistently compared with the baseline and achieved an average recognition increase of 15.60% by combining with MLLR speaker adaptation than MLLR alone. Such results proved the state-restructuring method could effectively reduce the recognition rate decreasing led by the difference between the decision tree structures of training data and testing data.