基于BiRNN的维吾尔语情感韵律短语注意力模型

Uyghur Sentiment Rhythm Phrase Attention Model Based on BiRNN

当前维吾尔语情感语音合成采用韵律边界预测方法来实现情感语音转换。通过该方法合成出来的语音,虽然可表现出相应的情感,然而其情感表现力不够理想。针对此问题,该文提出一种基于BiRNN的维吾尔语情感韵律短语注意力模型。在情感韵律转换前使用该模型进行情感分类,并将其分类结果作为韵律边界预测的输入,改进了情感韵律转换方法。使用改进的词性特征向量和韵律短语向量作为词向量的补充,从而有效提升维吾尔文文本情感分类的准确率。实验结果表明,该模型由两个单词构成的韵律短语作为特征时,准确率在维吾尔五分类情感数据集上达到了很好的分类效果。

At present,Uyghur sentimental speech synthesis uses prosodic boundary prediction method to realize emotional speech conversion.The speech synthesized by this method can express the corresponding emotions,but its emotional expression is not ideal.To solve this problem,this paper proposes an attention model of Uygur emotional prosodic phrases based on BiRNN.The model is used to classify emotion before prosodic conversion,and the classification results are used as input for prosodic boundary prediction to improve the method of prosodic conversion.The improved part-of-speech feature vector and prosodic phrase vectors are used to supplement the word vector,which effectively improve the accuracy of Uyghur text sentiment classification.The experimental results show that when the prosodic phrase composed of two words is used as a feature,the accuracy of the model achieves the best classification effect on the Uyghur five-category sentiment data set.