基于CNN-GRU-CTC的藏族学生普通话发音偏误检测  

CNN-GRU-CTC Based Detection ofPutonghua Mispronunciation By Tibetan Students

作　　者：梁青青 周小燕 赵春艳 LIANG Qing-qing;ZHOU Xiao-yan;ZHAO Chun-yan(School of Media Engineering,Lanzhou University of Arts and Science,Lanzhou 730000,China)

机构地区：[1]兰州文理学院传媒工程学院,甘肃兰州730000

出　　处：《兰州文理学院学报（自然科学版）》2023年第5期47-51,共5页Journal of Lanzhou University of Arts and Science(Natural Sciences)

基　　金：校级杰出青年科研人才培育计划项目(2018JCQN010);甘肃省高校教师创新基金项目(2023B-258)。

摘　　要：为了提高藏族学生学习普通话的发音水平,根据普通话和藏语发音特点设计并录制了一个偏误语音语料库,并结合卷积神经网络(Convolutional Neural Network,CNN)、门控循环单元(Gated Recurrent Unit,GRU)技术和连接时序分类技术(Connectionist Temporal Classification,CTC)搭建CNN-GRU-CTC声学模型,提出了一种发音偏误检测的方法.该方法将语音转换为一张图像作为输入,对完整的语谱图进行数据提取,利用深度全序列卷积神经网络进行建模,使用自动语音识别框架来进行发音偏误检测.实验结果表明:在该模型下,系统检测准确率为88.55%,错误拒绝率为7.16%,联合错误率为14.94%.该方法可以有效检测出错误发音,性能优于其他模型,可以用于检测和纠正藏族学生学习普通话时的错误发音,提高藏族学生的普通话发音水平.In order to improve the pronunciation level of Tibetan students learning Putonghua,this paper designs and records the error speech corpus according to the characteristics of Putonghua and Tibetan pronunciation.Combined with Convolutional Neural Network(CNN),Gated Recurrent Unit(GRU)and Connectionist Temporal Classification(CTC),a model of CNN-GRU-CTC is built,and a pronunciation error detection method is proposed.The method converts the speech into an image as input,extracts the data from the complete speech spectrum,uses the deep full sequence convolutional neural network for modeling,and uses the automatic speech recognition framework to detect the pronunciation bias.The experimental results show that the system detection accuracy rate is 88.55%,the false rejection rate is 7.16%,and the joint error rate is 14.94%under the model.The method can effectively detect mispronunciation.The performance of this model is superior to the results of the other models,which can be used to detect Tibetan students’mispronunciations of Putonghua learning and provide corrective feedback to help them improve their Putonghua pronunciation level.

关 键 词：发音偏误检测 卷积神经网络 门控循环单元 连接时序分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]