双阈值的特定英语音频句子边界检测  

Boundary detection of special English audio sentence based on dual-threshold

作　　者：刘秉权[1] 徐帅[1] 李相前[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《哈尔滨工业大学学报》2010年第2期259-263,共5页Journal of Harbin Institute of Technology

基　　金：国家自然科学基金资助项目(60673037);国家高技术研究发展计划资助项目(2006AA01Z197);黑龙江省自然科学基金资助项目(E200635)

摘　　要：为了提高英语音频句子切分的效果,提出了基于双阈值的句子边界检测方法.该方法针对VOA、BBC等特别适合英语学习者的音频所具有的波形规范、环境噪声小、速率通常比较稳定等特点,利用静音能量阈值和静音时延阈值来检测音频句子的边界,并辅以对照文本信息进行校正.针对VOA慢速英语的实验结果表明:单纯使用双阈值方法,音频切分的召回率超过96%,精确率超过94%;利用对照文本校正后,可进一步提高精确率.To improve the effect of sentence-evel English audio segmentation, a method of sentence boundary detection based on dual-threshold is proposed. With consideration of the characteristics of normative wave-form, small noise and stable speed for English audio such as VOA and BBC, the method in this paper detects the sentence boundary of English audio via its quiet energy threshold and quiet delay threshold, and the corre- sponding audio text is used to revise the segmentation result. Experiments on special English of VOA show that the recall rate of segmentation exceeds 96% and the precision rate exceeds 94% by using the dual-threshold method only. After revision via the corresponding text, the precision rate can be improved further.

关 键 词：音频切分 边界检测 双阈值 

分 类 号：TP37[自动化与计算机技术—计算机系统结构]