低信噪比环境下语音端点检测改进方法  

An improved speech endpoint detection method under low SNR

作　　者：王瑶 曾庆宁[1] 龙超[1] 谢先明[1] 毛维 WANG Yao;ZENG Qing-ning;LONG Chao;XIE Xian-ming;MAO Wei(Key Laboratory of Cognitive Radio and Information Processing of Ministry of Education,Guilin University of Electronic Technology,Guilin 541004,Guangxi,China)

机构地区：[1]桂林电子科技大学认知无线电与信息处理教育部重点实验室,广西桂林541004

出　　处：《声学技术》2018年第5期457-467,共11页Technical Acoustics

基　　金：国家自然科学基金(61461011);"认知无线电与信息处理"教育部重点实验室2016年主任基金(CRKL160107);广西自然科学基金重点项目(2016GXNSFDA380014)

摘　　要：针对语音端点检测在低信噪比环境下普遍存在检测性能急剧下降的问题,提出一种将调制域(时间-频率域)谱减法和自相关函数相结合的语音端点检测算法。该算法首先利用调制域谱减法较好的消噪能力来提高含噪语音的信噪比;然后根据语音和噪声的自相关函数的主峰最大值和次大值之比差异较大的特性,结合基于对数能量和自相关函数的端点检测方法对消噪后的语音进行端点检测。实验结果表明,该算法在低信噪比的环境下能取得较好的端点检测效果,并具有较好的稳健性。In this paper, a new approach combining the spectral subtraction in modulation （time-frequency）domain and the post processing for the autocorrelation functions of signal and noise is proposed to improve the performance of speech endpoint detection in low SNR environment. Firstly, the modified spectral subtraction used in modulation domain reduces the noise to increase SNR. Then, according to the feature we figure out that a quite difference in the ratio of maximum to secondary value of the peak of autocorrelation function exists between speech and noise, a method based on logarithmic energy and autocorrelation function is used for endpoint detection of the speech after denoising. Experiments show that the proposed method achieves a high performance and good robustness of speech endpoint detection under low SNR.

关 键 词：低信噪比 调制域 自相关函数 对数能量 端点检测 

分 类 号：TN912.35[电子电信—通信与信息系统]