【作者】 黄翔东； 蒋长丽； 王博；

【Author】 HUANG Xiangdong,JIANG Changli,WANG Bo School of Electronics and Information Engineering,Tianjin University,Tianjin 300072,China

【机构】 天津大学 电子信息工程学院；

【摘要】 为降低语音信号盲分离的计算复杂度及扩大算法的适用范围,提出基于稀疏时频特征模式提取的盲分离算法。该算法利用短时傅里叶变换的高度冗余性(保证了观测信号时频分布的稀疏性)及人耳的屏蔽效应,去掉能量值小的时频点以减少聚类所需的特征模式数量,然后通过时频屏蔽的方法完成语音信号的分离。实验结果表明,该算法所需特征模式数量仅为传统算法的5%～6%,加快了计算速度,节省了存储空间,因而算法是高效和可行的。更多还原

【Abstract】 In order to reduce the computation complexity of blind source separation of speech signals and expand the application scope of algorithm,this paper presents a blind source separation algorithm based on extracting the sparse time-frequency characteristic patterns.The small energy of the time-frequency s\ot(f,t )will be removed using the highly redundant of the STFT framework(ensuring the sparseness of T-F distribution of observation signals) and the masking effect of ears,to reduce the number of clustering patterns.Then it uses time-frequency masking to complete the separation of the speech signals.Experimental results show that the patterns only account for 5~ 6 percents of all patterns,which accelerate the calculation speed and save the storage space.Thus the high-efficiency and feasibility of the proposed method are verified.更多还原