基于子带粒子滤波的一种语音增强方法  

Subband particle filtering for speech enhancement

作　　者：金乃高[1] 殷福亮[1] 王冬霞[1] 陈喆[1] 

机构地区：[1]大连理工大学电子与信息工程学院,辽宁大连116024

出　　处：《通信学报》2006年第4期23-28,共6页Journal on Communications

基　　金：国家自然科学基金资助项目(60372082;60172073)~~

摘　　要：结合多采样率系统理论中的子带技术与贝叶斯估计理论中的粒子滤波技术,提出了一种基于子带粒子滤波的语音增强方法。该方法首先将语音信号分解成子带信号,建立各子带信号的低阶时变AR模型;然后利用R-B粒子滤波估计时变AR模型参数,对子带信号进行滤波处理;最后根据滤波后的子带信号重构语音信号,实现语音增强。该方法通过子带分解降低了R-B粒子滤波中采样空间的维数,在降低计算量的同时,提高了语音增强系统的性能。计算机仿真结果验证了该方法的有效性。Rao-Blackwellised particle filtering provided an effective solution for speech enhancement. A novel method was proposed for speech enhancement based on rao-blackwellised particle filters and subband decomposition. In rao-blackwellised particle filters, the dimension of sampling space was reduced by subband decomposition. Subband speech signals were obtained by subband decomposition, then modeled as low-order time-varying AR processeS. Rao-blackwellised particle filters were applied to estimate the parameters of AR model and filter the subband speech signals. The enhanced fullband speech signals were finally reconstructed by synthesizing the enhanced subband speech signals. Simulation results show that the proposed method not only has lower computational complexity, but also gives bet ter speech enhancement performance.

关 键 词：语音增强 R-B粒子滤波 子带分解 余弦调制滤波器组 

分 类 号：TN911[电子电信—通信与信息系统]