论文部分内容阅读
分别采用基于数据聚类和基于先验知识的两种研究方法,深入探讨了性别、口音、语速、信道等非语境因素对语音数据分类与建模的影响。为了综合考虑语境、非语境因素在统一框架下建模的问题,采用非语境因素扩展决策树方法。而对于这种方法生成的多套非语境因素相关的高精度声学模型,提出一种依据最大似然准则,动态组合生成测试人相关声学模型的算法。这种方法可以使系统相对误识率平均降低8％-10％。实验结果说明为非语境因素分类建模可以提高声学模型的建模能力,而且模型组合算法可以有效解决统一建模所带来的模型选择问题。
Two kinds of research methods based on data clustering and prior knowledge respectively are used to discuss the influence of non-context factors such as gender, accent, speech rate and channel on classification and modeling of speech data. In order to synthetically consider the contextual and non-contextual factors modeling under the unified framework, the non-contextual factors are used to extend the decision tree approach. As for the high-precision acoustic models related to non-contextual factors generated by this method, an algorithm based on the maximum likelihood criterion is proposed to dynamically generate test-related acoustic models. This method can reduce the relative error rate of the system by an average of 8% -10%. The experimental results show that the modeling of non-contextual factors can improve the modeling ability of acoustic model, and the model combination algorithm can effectively solve the problem of model selection brought by unified modeling.