基于机器学习的普通话韵律规则提取

EXTRACTING MANDARIN PROSODIC PATTERNS BY MACHINE LEARNING

韵律规则对于语音识别和语音合成研究具有重要意义 .目前的韵律规则大多是根据语言学的研究得出的一些定性的描述 .为了提取出更精确的定量描述的韵律规则 ,利用聚类分析提取出句子中音节的基频模式 ,在此基础上使用决策树进行韵律规则的学习 ,获得了较好的实验结果 .文中首先讨论韵律规则和聚类分析及决策树 ,然后给出数据预处理技术及所采用的学习算法 。

The pitch models play an important role in speech recognition and synthesis. Most models in using are extracted by linguistics experts, some of which are described qualitatively and of low precision. To acquire more accurate prosodic rules, clustering analysis and decision tree are employed to extract prosodic rules from actual speech, and the result is encouraging. This paper introduces prosodic rules, clustering analysis and decision tree firstly, then each stage of the process is discussed in detail, finally experiments are given.