基于规则学习的韵律结构预测

韵律结构的分析和预测作为提高语音合成系统自然度的一个重要核心组成, 日益受到重视。本文提出了一种基于规则学习的汉语韵律结构预测方法, 该方法从人工韵律标注的语料库中抽取语言学特征和两级韵律结构标记, 构建了实例数据库（example database), 再利用规则学习(rule learning)算法从实例中自动归纳韵律短语预测规则。本文通过大量的实验挑选出对于汉语韵律结构预测最有效的特征, 采用和比较了两种典型的规则学习算法。同时, 对于实验结果给出了较为系统的评价参数。实践表明, 规则学习算法用于韵律结构预侧达到了90%以上的正确率, 优于目前其他方法的结果, 是一种行之有效的办法。

Rule-learning Based Prosodic Structure Prediction

In this paper,a rule-learning based approach is proposed to predict prosodic structure from unrestricted Chinese text. Firstly, a speech corpus is collected, whose text is automatically segmented and tagged and further labeled with two-level prosodic structure and syntactic phrase boundaries. Secondly, features related to prosodic structure are extracted with the corresponding boundary types to establish an example database. Lastly, rule-learning algorithms are applied on the database to induce prediction rules by machine. Various experiments have been conducted to select the best features. Two typical learning algorithms(C4. 5 and Transformation-based learning)are experimented and compared with other methods. The paper also suggests general evaluation parameters for prosodic structure prediction. The experiments show that the rule-learning approach can achieve a better accuracy rate of 90% than the others. Thus it is justified as an effective way to prosodic structure prediction.