论文部分内容阅读
基于决策树理论的上下文相关声学模型在英语语音识别中已经得到了比较深入的研究和应用，但在汉语语音识别中的应用则研究的比较少。本文基于决策树理论建立了汉语语境相关模型-三音于模型，讨论了决策构建模所要解决的几个重要问题：（1）基本建模单元集的选择，（2）音子类别集的设计，（3）评估函数的选择，（4）停止准则的选择，（5）决策树的建立和三音子模型的生成，本文着重分析了两种不同建模单元的性能：对音子类别集的设计提出了一些一般性的准则，并对我们设计的类别集进行了统计分析；分析了三音子模型在语音库的覆盖程度。实验结果表明，基于决策树的三音子声学模型建立的识别系统与双音子声学模型系统比较，误识率下降了24．7％。
The context-dependent acoustic model based on decision tree theory has been deeply studied and applied in English speech recognition. However, the application of Chinese speech recognition is less studied. Based on the decision tree theory, this paper builds a Chinese context-related model - three tones in the model, discusses several important issues to be solved in decision modeling: (1) the selection of basic modeling unit sets, (2) (3) the choice of evaluation function, (4) the choice of stopping criterion, (5) the establishment of decision tree and the generation of triphone model. This paper focuses on the analysis of the performance of two different modeling units: The design of the set raises some general guidelines and makes a statistical analysis of the set of categories we designed. The degree of coverage of the three-tone model in the speech library is analyzed. The experimental results show that the recognition rate of the recognition system based on the triphone acoustic model based on decision tree is 24.7% lower than that of the two-tone acoustic model system.