一种新的基于频域独立成分分析的语音信号盲分离方法
A new blind speech signal separation method based on frequency domain independent component analysis
吴奇昌 1马峰 1戴礼荣1
作者信息
- 1. 中国科学技术大学电子工程与信息科学系科大讯飞语音实验室,安徽合肥230027
- 折叠
摘要
在频域利用传统的ICA进行分离时,如果分离矩阵没有经过良好的初始化,算法的收敛与分离性能都不够理想.本文提出了一种新的基于频域独立成分分析(ICA)的语音信号盲分离方法.首先通过分析混合信号的时频域特性对各个频带的分离矩阵进行初始化,使算法的收敛速度更快,并很好的解决了输出信号的次序不确定性问题;进一步根据以初始化的分离矩阵分离出的源信号间的幅度相关性,仅挑选出一部分频带进行ICA的迭代,最终达到在追求良好分离性能的同时极大提升运算效率的目的.仿真的无回声环境和几种实际的回声环境下所得到的实验结果表明,该方法在分离性能和算法效率上均优于传统的频域ICA方法.
Abstract
In conventional FDICA method,convergence and separation performance degrade if the separation matrix is not well initialized.This paper presents a novel blind source separation method,which is based on frequency-domain independent component analysis (FDICA).First,the separation matrix of each frequency bin is initialized through analyzing the characteristic of the mixed signals in time-frequency domain; it contributes to the faster convergence,and solves the ambiguity of the frequency permutation.Further,by using amplitude correlation among the source signals separated by the initialized separation matrix,a few frequency bins are selected to perform ICA,significantly to improve computational efficiency with good separation performance.Experimental results in anechoic and several reverberant conditions show that the proposed method outperforms the conventional FDICA in both separation performance and computational efficiency.
关键词
盲信号分离/独立成分分析/时频分析/卷积混合/频带挑选Key words
blind signal separation/ICA/time-frequency analysis/convolutive mixing/band selection引用本文复制引用
出版年
2013
NETL
NSTL
维普
万方数据