结合分层条件随机场与标点符号的维吾尔语韵律边界预测

Uyghur Language Prosodic Boundary Prediction Combined with Hierarchical Conditional Random Field and Punctuation

韵律结构的正确预测是高自然度语音合成系统的重要组成部分。针对维吾尔语的黏着性特点,给出其相应的韵律层次结构,采用基于条件随机场(CRF)的分层自底向上方法预测维吾尔语的韵律词和韵律短语边界,并将维吾尔语形态特征作为韵律边界预测模型的重要特征。为进一步纠正韵律边界预测错误并消除标点符号边界中不同韵律边界之间的歧义,以标点符号边界为单位建立基于CRF的标点符号韵律边界预测模型,并与双层自底向上CRF模型相结合,提出一种韵律边界预测方法。通过对不同的特征模板和模型进行反复实验,以得到更好的韵律边界预测性能。实验结果表明,该方法明显提高了韵律边界的预测召回率。

Correct prosodic boundary prediction is crucial for the quality of synthesized speech. This paper presents the prosodic hierarchy of Uyghur language which belongs to agglutinative language. A two-layer bottom-up hierarchical approach based on Conditional Random Field （CRF） is used for predicting prosodic word and prosodic phrase boundaries. Morphological features are considered useful for prosodic boundary prediction and added into the feature sets. In order to further enhance the accuracy of prosodic boundary prediction at punctuation sites, CRF based prosodic boundary determination method is used and integrated with bottom-up hierarchical approach. Consequently, the best prosodic boundary prediction performance is achieved by large and repeated experiment of different feature sets and different models. Experimental results show that the proposed method obviously improves the recall rate prediction of the prosodic boundary.