基于CNN-HMM和RNN的维吾尔语语音识别  

Uyghur speech recognition based on CNN-HMM and RNN

作　　者：穆凯代姆罕·伊敏江 艾斯卡尔·艾木都拉[1] 米吉提·阿不里米提[1] IMINJAN Mukaddam;HAMDULLA Askar;ABLIMIT Mijit(College of Information Science and Engineering,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《现代电子技术》2021年第11期172-176,共5页Modern Electronics Technique

基　　金：国家自然科学基金项目(61662078);国家重点研发计划(2017YFC0820602)。

摘　　要：神经网络模型的发展给资源匮乏语言的语音及语言信息处理带来新的机遇,基于神经网络的少数民族语言的语音识别系统效率及准确率比传统方法有了很大提高。对于大词汇量语音识别系统,适当选择声学模型和语言模型很重要。对较小的维吾尔语语料库(THUYG公开语料库)进行了深入研究,采用Kaldi开源语音识别平台将深度的CNN-HMM作为声学模型,通过理论分析和对比实验,分别在N-gram和RNN两种语言模型上进行对比实验。实验结果表明,基于神经网络RNN语言模型的系统有更好的识别效果,提升了维吾尔语语音识别准确率,并将词错误率降到15.06%。The development of neural network model brings new opportunities for the speech and language information processing of languages with scarce resources.The efficiency and accuracy of the minority language speech recognition system based on neural network have been greatly improved compared with the traditional methods.For speech recognition systems with large vocabularies,it is important to adopt appropriate acoustic model and language model.In this paper,an in-depth research on the smaller openly available Uyghur corpus(THUYG open corpus)is carried out.In this research,the Kaldi open source speech recognition platform is used to build a CNN(convolutional neural network)-HMM(hidden Markov model)based acoustic model.Comparative experiments were performed on the N-gram and RNN(recurrent neural network)language models.Experimental results show that the speech recognition of CNN-HMM acoustic model combined with RNN language model show better achievements and the word error rate is reduced to 15.06%.

关 键 词：语音识别 维吾尔语 声学模型 语言模型 CNN-HMM N-GRAM语言模型 循环神经网络 Kaldi 

分 类 号：TN912-34[电子电信—通信与信息系统]