基于深度神经网络的资源匮乏语言语音关键词检索  

Resource deficient languages′speech keyword retrieval based on deep neural network

作　　者：张伟涛 米吉提·阿不里米提[1] 郑方 艾斯卡尔·艾木都拉[1] ZHANG Weitao;Mijit Ablimit;ZHENG Fang;Askar Hamdulla(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》2022年第11期68-72,共5页Modern Electronics Technique

基　　金：国家重点研发计划(2017YFC0820602)。

摘　　要：资源匮乏语言语音信息检索研究比汉语、英语等大语言进展缓慢,需要大量预处理工作。神经网络模型在低资源环境下的高效建模能力给低资源语言信息处理工作带来便利。文中以维⁃哈等低资源语言为基础,通过一系列预处理过程获得了这些语言的语音及文本资源,再利用高斯混合隐马尔可夫模型GMM⁃HMM、深度神经网络隐马尔可夫模型DNN⁃HMM等完成了关键词检索实验。实验结果表明,三音素下的DNN⁃HMM模型比GMM⁃HMM模型检索性能要好。维吾尔语的ATWV达到了0.368,MTWV达到了0.491,检索结果准确率达到了89.36%;哈萨克语的ATWV达到了0.382,MTWV达到了0.421,检索结果准确率达到了82.15%。Researches on the speech information retrieval of resource deficient languages progress more slowly than that of the English,the Chinese and the other major languages because of lots of pre⁃processing.The efficient modeling ability of neural network models in low resource environment brings opportunity to information processing of the minority languages.In this paper,the minority languages,for example,the Uyghur and the Kazakh,are taken as the basis.The phonetic and text resources of these languages are obtained by a series of pre⁃processing.On the basis of the above,the GMM⁃HMM(Gaussian mixture model⁃hidden Markov model)and DNN⁃HMM(deep neural network⁃hidden Markov model)are utilized to accomplish the experiment on keyword retrieval.The experimental results show that the retrieval performance of triphone model DNN⁃HMM is better than that of the triphone model GMM⁃HMM.The Uyghur′s ATWV(actual term⁃weighted value)can reach 0.368,its MTWV(maximum term⁃weighted value)can reach 0.491,and its retrieval accuracy can reach 89.36%.The Kazakh′s ATWV can reach 0.382,its MTWV can reach 0.421,and its retrieval accuracy can reach 82.15%.

关 键 词：语音关键词检索 维吾尔语 哈萨克语 深度神经网络 检索流程 声学模型 

分 类 号：TN711-34[电子电信—电路与系统] TP391.1[自动化与计算机技术—计算机应用技术]