基于方差和谱熵结合的语音端点检测方法  

A Speech Endpoint Detection Method Based on Variance and Spectral Entropy

作　　者：毛强 刁鸿鹄 章小兵 MAO Qiang;DIAO Honghu;ZHANG Xiaobing(School of Electrical and Information Engineering, Anhui University of Technology, Maanshan 243032)

机构地区：[1]安徽工业大学电气与信息工程学院,安徽马鞍山243032

出　　处：《常州工学院学报》2021年第2期36-40,52,共6页Journal of Changzhou Institute of Technology

摘　　要：针对传统方法在低信噪比条件下的检测结果无法较好地满足需要,提出了将语音信号小波包BARK子带方差和谱熵二者相结合的新端点检测方法。首先对带噪语音采用多窗谱估计谱减法降噪,然后利用小波包分解构成BARK子带,求出每帧信号的BARK子带方差均值和谱熵值,最后用方差值除以谱熵值,将二者的比值作为双门限检测法的参数进行端点检测。实验证明,在低信噪比条件下,该方法相比于传统方法的检测效果更佳。In terms of the problem that the detection results of traditional methods in low SNR condition can not well meet the needs,a new endpoint detection method that combines the variance and spectral entropy of voice signal wavelet packet BARK subband is proposed.First,use the spectral subtraction of multi-window spectrum estimation to denoise the noisy speech.Then,use wavelet packet decomposition to form BARK subband,calculating the mean variance and spectral entropy of BARK subband of each frame signal.Finally divide the variance value by the spectral entropy,and the ratio of the two is used as the parameter of the double threshold detection method for endpoint detection.The experiment proves that under the low SNR condition,this method has better detection effect than traditional methods.

关 键 词：多窗谱估计谱减 小波包 BARK子带方差 谱熵 端点检测 

分 类 号：TN912.3[电子电信—通信与信息系统]