基于多分布深度神经网络英文重音检测方法

English word stress detection method based on multi-distributed deep neural network

为了利用英语语音进行发音错误检测和诊断,研究了超音段中的英语词汇重音和音调重音检测,使用了多分布神经网络方法。多分布神经网络的隐藏单元和可见单元均为二进制,利用基于音节的韵律特征和规范的词汇重音模式作为输入特征。将该方法与已有方法进行了比较实验,结果表明:基于MD-DNN的词汇重音检测方法对三个或三个以上音节的单词进行音节分类(主/次/无重音)的准确率为87.9%,性能明显优于使用高斯混合模型(GMM)和显著性模型(PM)。与词汇重音检测方法相似,音高重音检测的准确率为90.2%,分别比GMM和PM检测的准确率高9.6%和6.9%。

In order to detect and diagnose pronunciation errors by using English speech,this paper studies the detection of English lexical and tonal stress in suprasegmental speech by using the multi-distributed neural network method.The hidden and visible units of the multi-distributed neural network are binary,and the syllable based prosodic features and standardized lexical stress patterns are used as input features.The proposed method is compared with the existing methods.The experiment results show that the accuracy of the MD-DNN method is 87.9%in the classification of syllables(primary/secondary/unstressed)for words with three or more syllables,which is significantly better than the Gaussian Mixture Model(GMM)and the Significance Model(PM).Similar to the lexical stress detection method,the accuracy of pitch stress detection is 90.2%,which is 9.6%and 6.9%higher than that of GMM and PM,respectively.