藏语孤立词语音识别技术研究  

Research on Tibetan isolated word speech recognition technology

作　　者：赵尔平[1] 王聪华[1] 党红恩[1] 雒伟群[1] 

机构地区：[1]西藏民族大学信息工程学院,陕西咸阳712082

出　　处：《西北师范大学学报（自然科学版）》2015年第5期50-54,共5页Journal of Northwest Normal University(Natural Science)

基　　金：国家自然科学基金资助项目(61162025);西藏自治区自然科学基金资助项目(12KJZRYMY07);西藏自治区科技厅重点项目(藏科发[2013]189号);西藏民族学院重大科研项目(11myZ05)

摘　　要：针对藏语读音首先看后加字,然后根据元音的位置关系决定读音,而且元音比辅音携带更多听觉感知信息的特点,提出了一种改进的HTK系统藏语孤立词语音识别技术.在识别特征参数中,增加更能表征元音特征的共振峰参数提高语音识别的正确性,通过循环迭代方法提高语音训练速度,利用藏文字母拉丁转写方法解决藏文和语音识别系统编码不一致的问题.在二次开发的HTK平台进行实验,正确率达到92.83%,实验结果表明元音特征在藏语音识别中起到重要作用.Aiming at Tibetan pronunciation firstly look after hong jia zi , then its pronunciation is determined by the position of vowel , and a vowel carry more auditory perception information than a consonant in speech ,a Tibetan isolated word speech recognition technology of improved HTK system is proposed in this paper . The accuracy of speech recognition is improved by increasing a formant parameter in the recognition characteristic parameters , the formant parameter can characterize vowel features very well , the speech training speed is raised by cycle iteration , Tibetan letters transformation Latin alphabet solves inconsistent problem that Tibetan and speech recognition system code . The test is executed on the secondary developing HTK platform , the correct rate reaches 92.83% . Experimental result indicates that vowel features play an important role in Tibetan speech recognition .

关 键 词：藏语孤立词 共振峰 M el倒谱特征 循环迭代 隐马尔可夫模型 语音识别 

分 类 号：TP912.34[自动化与计算机技术]