基于短时能零熵的端点检测方法  

Endpoint Detection Algorithm Based on Short-term Energy-zero Spectral Entropy

作　　者：尹岩岩[1,2] 殷业[1] 罗汉文[3] 钱栋军[1] 

机构地区：[1]上海师范大学信息与机电工程学院,上海200234 [2]上海真灼电子技术有限公司,上海201100 [3]上海交通大学电子工程系,上海200030

出　　处：《计算机仿真》2012年第11期408-411,共4页Computer Simulation

基　　金：科研计划专项上海市科委支持项目(10dz1110900);国际技术转移上海市科委支持项目(10510708400)

摘　　要：端点检测技术是语音识别系统预处理阶段中的第一个关键技术,而传统的端点检测特征参数LPC距离、倒谱特征、TF参数和分形特征等参数的运算量太大,对硬件要求很高,阻碍了人机交互技术在日常生活中的普及。通过对语音信号三个端点检测的特征参数短时平均过零率、短时能量和基本谱熵逐一分析研究,提出了一种新型的语音参数——短时能零熵值,并作为主要参数应用于端点检测中。实验证明,短时能零熵值结合了信号的时域和频域语音特征,能够对背景噪声做出反馈、并且可以在复杂的背景噪音环境下对语音和非语音做出有效、稳定的区分,其端点检测的隔离度较大,具有稳健的抗噪特性。Speech endpoint detection is the first key technology encountered in pre-processing stage of voice rec- ognition system. The traditional characteristic parameters of endpoint detection are too complex and demanding hard- ware, it is not conducive to make interactive technology more popular in daily life. Three basic characteristic parame- ters of endpoint detection short-time average zero crossing rate, short-term energy and spectral entropy were analyzed separately in this treatise. Based on these three impact factors, and the Endpoint Detection Algorithm of the Short- term EZSE （Energy-zero Spectral Entropy） was proposed. EZSE combined the advantages of time-domain and fre- quency-domain phonetic features, which can make a feedback on the background noise and make effective distinction between the speech signals and non-voice signals under the environment of the complex background noise. And the i- solation of Endpoint detection is larger, with robust anti-noise characteristics.

关 键 词：端点检测 过零率 短时能量 谱熵 短时能零熵 

分 类 号：TN912.34[电子电信—通信与信息系统]