基于信噪比估计的说话人识别前端处理  

Speaker Recognition Front-end Processing Based on SNR Estimation

作　　者：尹海明[1] 王金明[2] 李欢欢[3] 

机构地区：[1]解放军理工大学通信工程学院研究生1队,江苏南京210007 [2]解放军理工大学通信工程学院 [3]解放军理工大学通信工程学院研究生5队

出　　处：《军事通信技术》2014年第4期43-48,共6页Journal of Military Communications Technology

基　　金：中兴通讯产学研合作研究资助项目(CON1307160001)

摘　　要：针对说话人识别系统的噪声鲁棒性问题,文章采用了基于信噪比估计从而选取高信噪比语音帧的前端处理方法。根据一定阈值去除语音中信噪比较低的语音帧,对保留的语音帧提取特征参数并进行识别。该方法的有效性取决于分帧信噪比估计的准确性。由于传统的谱减法以及滤波法难以对非平稳噪声信噪比进行准确估计,文中提出了使用改进的最小值控制递归平均算法进行信噪比估计来实现高信噪比帧筛选,实验结果表明,与基于维纳滤波语音增强的GMM-UBM系统对比,5dB street噪声下识别率由78.5%提升至85.5%,5dB car噪声下识别率由88%提升至91%。To improve the noise robustness of speaker recognition systems,the signal to noise ratio estimation front-end method was applied,discarding some speech frames of low SNR according to a certain threshold and using the remaining for feature extraction and recognition.The effectiveness of this method largely depends on the accuracy of the signal to noise ratio estimation.Traditional spectral subtraction or signal filtering cannot estimate the SNR of none stationary noisy speech accurately,so improved minima controlled recursive averaging algorithm was proposed instead.Compared with the wiener filtering speech enhancement based front-end system,experimental results show that in the none stationary 5dB street noise environment,the recognition rate is improved from 78.5%to 85.5%,and that in the stationary 5dB car noise environment,the recognition rate is improved from 88%to 91%.

关 键 词：说话人识别 信噪比估计 最小值控制递归平均 高斯混合模型 

分 类 号：TN912.34[电子电信—通信与信息系统]