基于DTW的俄语短指令语音识别  

Speech recognition of Russian short instructions based on DTW

作　　者：王彤 马延周[1] 易绵竹[1] 

机构地区：[1]中国人民解放军外国语学院语言工程系,河南洛阳471000

出　　处：《山东大学学报（理学版）》2017年第11期29-36,共8页Journal of Shandong University(Natural Science)

基　　金：国家自然科学基金重大项目(11590771)

摘　　要：面向训练语料有限的语音识别任务,基于动态时间规整(dynamic time warping,DTW)算法对俄语语音进行识别。首先,以跨语言标注的语音语料为资源基础,研究融合音字转换和机器翻译的语音识别方法。其次,结合俄语语音特点,以元音为中心设置动态门限阈值,实现精确至音节的端点检测,识别速度提高了34.4%,准确率提高了14%。然后,综合时域、频域分析,提取反映语音静态特征和动态变化的参数模板。另外,引入全局限制和早弃策略改进DTW算法,避免病态匹配,缩小计算规模,使速度提高了19.7%,准确率提高了4.8%。在俄语短指令语音集上做五折交叉验证,识别准确率达到74.9%。Focus on speech recognition task with limited training corpus,this paper makes research of Russian speech recognition based on DTW（ dynamic time warping） algorithm. Firstly,we study methods for combining speech recognition and machine translation with the speech corpus which annotating tags of cross language text. Secondly,based on the characteristics of Russian speech,in order to detected syllable endpoint,we set dynamic threshold according to the central vowel,which increased the speed by 34. 4% and increased the accuracy by 14%. Finally,we extract the parameters of the static and dynamic characteristics by analyzing speech features of time domain and frequency domain. In addition,the DTW algorithm is improved to overcome the ill condition and reduce the computation scale with global restrictions and early discard strategies,which increased the speed by 4. 8% and increased the accuracy by 19. 7%. Experiments on the Russian short instruction set with 5 fold cross validation,and the accuracy of speech recognition reached 74. 9%.

关 键 词：俄语语音识别 端点检测 DTW算法 跨语言语音识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]