一种基于多属性模糊决策的英文韵律短语边界预测方法

A Boundary Prediction Method for English Prosody Phrase Based on Multi-attribute Fuzzy Decision Making

对于各类TTS(Text to Speech)系统而言,能否准确地预测韵律短语边界对TTS系统的效果有着关键性的影响.目前常使用决策树来做韵律短语边界预测,但这种方法受到了训练数据的均衡性以及决策算法本身无法达到全局最优的制约.为了改善预测效果,在传统的决策树方法之上,将决策树使用的聚类属性与模糊决策相结合,提出通过多属性模糊决策方法来预测英文韵律短语边界.实验表明,使用这种方法后,效果比基于决策树的预测方法的效果有较大提升,F-Score由64. 4%提升到69. 3%,不可接受率也从28. 6%降低到21. 4%.

For all kinds of TTS(Text to Speech) systems,whether the prosodic phrase boundaries can be accurately predicted has a crucial impact on the TTS system. At present,decision tree is often used to predict the boundaries of prosodic phrases,but this method is restricted by the balance of training data and the decision algorithm itself can not achieve global optimum.Therefore,in order to improve the prediction effect,a multi-attribute fuzzy decision-making method to predict the boundaries of English prosodic phrases based on the traditional decision tree method is proposed in this paper,which combines the clustering attributes used in the decision tree with fuzzy decision-making. Experiments show that the effect of this method is better than that of the decision tree-based prediction method. The F-Score is increased from 64. 4%to 69. 3%,and the unacceptable rate is reduced from 28. 6% to 21. 4%.