基于声源方位信息和非线性时频掩蔽的语音盲提取算法  

Speech blind extraction algorithm based on sound source azimuth information and nonlinear time-frequency masking

作　　者：夏秀渝[1] 何培宇[1] 

机构地区：[1]四川大学电子信息学院,成都610064

出　　处：《声学学报》2013年第2期224-230,共7页Acta Acustica

基　　金：国家自然科学基金(61071159)资助项目

摘　　要：针对欠定卷积混合的语音信号模型,提出一种基于声源方位信息和非线性时频掩蔽的语音盲提取算法。首先对低频段混合语音信号进行时频分析估计瞬时相对时延(ITD)并采用势函数聚类分析方法估计出声源个数及其ITD,接着锁定目标提取准确的目标语音方位信息,最后利用独立语音在时频域上的近似W一分离正交性,采用非线性时频掩蔽的方法提取目标语音。仿真实验表明,该方法能锁定任意感兴趣目标方位,能有效提取目标语音,文中实验条件下信噪比增益平均达9.5 dB。For the underdetermined convolution mixture model, a new speech blind extraction algorithm based on sound source azimuth information and nonlinear time-frequency masking was proposed. At first, instantaneous ITDs were calculated through time-frequency analysis in lower frequency domain, and the number of sources and their ITDs were estimated using the potential function. Then the object source was locked and accurate azimuth information of object was estimated. At last, the object speech was extracted via nonlinear time-frequency masking which was based on the azimuth information of object. Simulation results showed that our proposed speech extraction algorithm can lock interested object speech from random direction and extract object speech effectively, the signal-noise-ratio gain （SNRG） was obtained 9.5 dB averagely in our experiment condition.

关 键 词：盲提取算法 语音信号 时频分析 方位信息 非线性 掩蔽 声源 聚类分析方法 

分 类 号：TN912.3[电子电信—通信与信息系统]