适用于短波信号侦察的话音端点检测方法  

Speech Endpoint Detection Method Applied to Shortwave Signal Reconnaissance

作　　者：张洪德 韩鑫怡 ZHANG Hongde;HAN Xinyi(Communications NCO Academy,Army Engineering University of PLA,Chongqing 400035,China)

机构地区：[1]陆军工程大学通信士官学校,重庆400035

出　　处：《陆军工程大学学报》2023年第1期63-70,共8页Journal of Army Engineering University of PLA

基　　金：军内科研项目(LJ20191C070659)。

摘　　要：针对传统话音端点检测方法在短波低信噪比信道下检测准确率低的问题,提出一种将深度生成对抗网络和自适应参数的子带对数能熵积相结合的话音端点检测方法。该方法首先利用深度生成对抗网络话音增强方法降低噪声对待检测信号的影响,再以自适应参数的子带对数能熵积这一新的话音特征参数为阈值,使用自适应阈值双门限检测法完成话音端点检测。仿真实验结果表明,该方法对于-5 dB信噪比的标准话音库检测的平均加权错误测度仅为13.5%,而对于实际短波侦察信号库检测的平均加权错误测度为16.7%,均优于能零熵法和多窗谱估计谱减与能熵积法。Aiming at the problem of low detection accuracy of traditional speech endpoint detection methods in short-wave and low signal-to-noise ratio channels,a speech endpoint detection method that combines deep generative adversarial networks and sub-band logarithmic energy entropic product of adaptive parameters is proposed.The method uses the deep generative adversarial network speech enhancement method to reduce the influence of noise on the signals to be detected.Then,taking the new speech characteristic parameter of the sub-band logarithmic energy entropic product of adaptive parameters as the threshold,this method uses the adaptive double-threshold to complete the speech endpoint detection.The simulation results show that the average weighted error measure of this method is only 13.5%for the standard speech library detection with-5 dB SNR,while the average weighted error measure for the actual shortwave reconnaissance signal library detection is 16.7%,and both are better than the energy-zero entropy method and the multi-window spectral estimation spectral subtraction and the energy entropy product methods.

关 键 词：深度生成对抗网络 话音增强 话音端点检测 对数能量 谱熵 

分 类 号：TN912.3[电子电信—通信与信息系统]