基于声母和深度哈希的密文语音全文检索方法  

Full-text retrieval algorithm for encrypted speech based on initial consonant and deep hashing

作　　者：胡颖杰 张秋余[1] 李昱州 HU Yingjie;ZHANG Qiuyu;LI Yuzhou(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学计算机与通信学院,甘肃兰州730050

出　　处：《华中科技大学学报（自然科学版）》2021年第12期83-88,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61862041).

摘　　要：为了实现基于内容的语音全文检索,提高语音检索性能,以及保障云端语音数据的隐私安全,提出了一种基于声母和深度哈希的密文语音全文检索方法.该方法将提出的基于汉语声母和元音的双向循环递归神经网络(RNN)-长短时记忆(LSTM)深度学习模型与语音感知哈希相结合,分别将加密语音和生成的哈希码上传至云端密文语音库和全文哈希索引表,并建立一一映射关系.查询时提取待查询语音的哈希码,并与云端的全文哈希索引表进行阶段式匹配检索.实验结果表明:该方法既能保障语音隐私安全,又能获得较高的检索精确度与可观的召回率(当精确度为97.68%时召回率可达47.60%),并在一定程度上减弱了说话人声音特征对全文检索的不利影响.To improve the retrieval performance and guarantee the privacy security of cloud speech data,a new encrypted speech retrieval algorithm was proposed based on initial consonant and deep hashing,aimed at the content based speech full-text retrieval.The proposed bidirectional recurrent neural network(RNN)-long short term memory(LSTM)model,which was on initial consonant and vowels of Chinese Pinyin,was used to generate perceptual deep hashing.Then the encrypted speech data and its hash code would be uploaded to cloud encrypted speech library and full-text hash index,and one to one mapping relationship would be built.To obtain the retrieval results,the hash code of query speech would be matched with the full-text hash index stage by stage.The experiments show that the proposed algorithm has high retrieval precision(97.68%)and considerable recall rate(47.60%)with good privacy security,to some extent it can weaken the negative impact of speaker's identity on full-text retrieval.

关 键 词：密文语音检索 全文检索 声母 深度学习 哈希编码 递归神经网络 长短时记忆 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] TN912.34[自动化与计算机技术—计算机科学与技术]