基于语音能量比的解决频域ICA次序不确定性问题的算法  

An algorithm for solving the permutation indeterminacy problem of frequency-domain ICA based on speech energy ratio

作　　者：王志强[1] 王涛[1] 金志文 WANG Zhiqiang;WANG Tao;JIN Zhiwen(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China;Unit 96216 of PLA,Beijing 100085,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]中国人民解放军93216部队,北京100085

出　　处：《上海大学学报（自然科学版）》2022年第2期226-237,共12页Journal of Shanghai University:Natural Science Edition

基　　金：国家自然科学基金资助项目(61671011)。

摘　　要：随着人工智能物联网(artificial intelligence&internet of things,AIoT)的发展,硬件技术的飞速进步,更多的智能音箱进入人们的生活,人机交互方式也从早期的遥控变成了人声控制.但设备中麦克风采集到的语音信号往往含有大量噪声和干扰人声,为此需对麦克风采集到的语音进行语音分离处理.常用的技术有频域独立成分分析(independent component analysis,ICA),但是频域ICA存在次序不确定性问题,即将分离出的源1分量分类到源2通道,将分离出的源2分量分类到源1通道,从而导致分离性能大大降低.为此,提出一种基于语音能量比来解决频域ICA中次序不确定性问题的算法,有效地提高了分离性能.在SiSEC(Signal Separation Evaluation Campaign)、ChiME(Challenge for Computational Hearing in Multisoure Environments)数据集上对分离性能进行实验,所得结果比已有算法均有提升,且针对强混响环境下的混合信号依然保持良好的分离性能.With the development of artificial intelligence&internet of things(AIoT)and the rapid advancement of hardware technology,an increasing number of smart speakers are becoming a part of people’s lives.Human-computer interaction has also witnessed a shift from remote control to voice control.However,the audio signals recorded by the microphone in a device usually contain considerable noise and interfering voices.Therefore,separation needs to be performed on the signals recorded by the microphones.Frequencydomain independent component analysis(ICA)is a commonly used separation technique,but it faces the permutation indeterminacy problem,i.e.,the separated components from Source 1 are classified into a channel for Source 2,whereas the separated components from Source 2 are classified into a channel for Source 1,which greatly deteriorates the separation performance.To address this issue,we proposed an algorithm based on the speech energy ratio,which effectively improved the separation performance.The separation performance was tested on the Signal Separation Evaluation Campaign(SiSEC)and Computational Hearing in Multisource Environments(CHiME)datasets.The results showed that the proposed algorithm outperformed existing algorithms,and a good separation performance for mixed signals could be maintained even in an environment with strong reverberations.

关 键 词：盲源分离 语音分离 频域独立成分分析 次序不确定性 能量比 

分 类 号：TN912.3[电子电信—通信与信息系统]