基于多频段能量相关排序的语音卷积混合盲源分离  

Blind convolution speech separation based on multi-band ordering

作　　者：欧旭东[1] 张天骐[1] 闫振华[1] 张世会[1] 

机构地区：[1]重庆邮电大学信号与信息处理重庆市重点实验室,重庆400065

出　　处：《计算机应用研究》2016年第5期1481-1485,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61371164;61275099;61102031);信号与信息处理重庆市市级重点实验室建设项目(CSTC2009CA2003);重庆市杰出青年基金资助项目(CSTC2011jjj140002);重庆市自然科学基金资助项目(CSTC2012JJA40008);重庆市教育委员会科研项目(KJ120525;KJ130524);重庆市研究生科研创新项目(CYS14140)

摘　　要：针对语音卷积盲源分离频域法排列顺序不确定性问题,提出一种多频段能量排序算法。通过对混合信号的短时傅里叶变换(STFT),在频域上各个频点建立一个瞬时混合模型进行独立分量分析,之后结合能量相关排序法和波达方向(DOA)排序法解决排序不确定性问题,再利用分裂语谱方法解决幅度不确定性问题,进而得到每个频点正确的分离子信号,最后利用逆短时傅里叶(ISTFT)变换得到分离的源信号。仿真结果表明,与Murata的排序算法对比,改进的算法在信号偏差比、信道干扰比、系统误差比上都所提高。For the ranking uncertainty in frequency domain of speech convolutive blind source separation,this paper developed a new method based on multi-band energy sorting algorithm. Firstly,through the short Fourier transform（ STFT） for the mixed signals,it built an instantanneous mixing model in frequency domain of each point and then used independent component analysis to separate it. After that,it proposed a multi-band energy sorting algorithm which was based on combining energy-related and the direction of arrival（ DOA） sorting methods to solve the problem of ranking uncertainty. Then it used the split speech spectral method to solve the problem of the uncertainty of magnitude,and got the proper sub-signals of each frequency. Finally,it separated out the source singals through the inverse short Fourier tranform. The simulation experiments show that the proposed method has better source to distortion,source to interference ratio and source to artifacts ratio.

关 键 词：卷积盲源分离 短时傅里叶变换 分裂语谱 波达方向 

分 类 号：TN911.7[电子电信—通信与信息系统]