分层特征模板筛选的维吾尔语韵律边界预测

Uyghur prosodic boundary prediction based on hierarchical feature template selection

韵律边界的正确预测是连续语音合成系统中首要解决的关键问题。针对维吾尔语分层韵律结构,通过基于条件随机场(CRF)的分层自底向上方法预测了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。根据不同韵律边界层次的特点,对分层韵律边界预测方法进行了改进,针对分层方法的不同层次进行独立特征模板筛选,从而进一步提高韵律边界预测性能。通过对不同的特征模板和不同韵律边界预测模型进行反复实验,得到了最好的预测性能。实验结果表明,该方法明显提高了韵律边界预测结果。

Accurate prediction of prosodic boundary is a very important task in an arbitrary Text-To-Speech(TTS)system.A two-layer bottom-up hierarchical approach based on conditional random fields is used to predict prosodic word and prosodic phrase boundaries according to the prosodic hierarchy of Uyghur language.Morphological features are also integrated into the feature sets which are used to train the prosodic boundary prediction model.In order to further enhance the accuracy of prosodic boundary prediction,a two-layer bottom-up hierarchical approach is improved by selecting independent feature sets at every layer according to the different prosodic boundary structure of corresponding layer.Consequently,the best prosodic boundary prediction performance is achieved by extensive and repeated experiment with different feature sets and different models.Experimental results show that the proposed method obviously improves the prediction of prosodic boundary.