基于对数能量倒谱特征的端点检测算法  

Voice activity detection using logarithmic energy and cepstrum Distance

作　　者：王民[1] 孙广[1] 沈利荣[2] 刘利[1] 

机构地区：[1]西安建筑科技大学信息与控制工程学院,西安710055 [2]西安石油大学光电油气测井与检测教育部重点实验室,西安710065

出　　处：《计算机工程与应用》2014年第16期198-201,215,共5页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61073196)

摘　　要：端点检测技术是语音识别的关键技术之一,为了克服传统倒谱距离语音端点检测算法在低信噪比下检测效果的不理想,将对数能量(LE)特征和倒谱(C)特征相结合,提出了一种新的对数能量倒谱特征(LEC),采用模糊C均值聚类和贝叶斯信息准则(BIC)方法估计特征门限,得出了正确的语音端点判断,在三种典型噪声下,对信噪比从-5 dB到15 dB的带噪声语音进行仿真,结果表明LEC法的检测错误率仅为20.25%,明显低于倒谱法和对数能量法,能有效地确定语音的端点并改善语音识别效果。Endpoint detection is one of the key technologies of speech recognition, in order to overcome the undesirable detection results of traditional cepstrum distance in speech endpoint detection algorithm under low signal to noise ratio, combined logarithmic energy feature（LE）with cepstrum features（C）for endpoint detection, proposes a new logarithmic energy cepstrum features（LEC）, uses fuzzy C-means clustering and Bayesian information criterion to estimate features threshold, achieves better endpoint judgment, conducts the SNR simulation from-5 dB to 15 dB with noisy speech under three kinds of typical noise. The results indicate that the LEC method’detection error rates is just 20.25%and significant-ly lower than cepstrum and logarithmic energy method, it also can effectively determine the speech endpoint and improve voice recognition results.

关 键 词：对数能量 倒谱距离 模糊C均值聚类 贝叶斯信息准则(BIC) 端点检测 Bayesian Information Criterions(BIC) 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]