【作者】 李鸿燕； 张雪英； 屈俊玲；

【Author】 Li Hongyan;Zhang Xueying;Qu Junling;College of Information Engineering, Taiyuan University of Technology;

【机构】 太原理工大学信息工程学院；

【摘要】 针对计算听觉场景分析传统分段计算量大、易受干扰的问题,提出了一种基于信号开始和截止时刻(onset/offset)的语音分离算法,将属于同一声源的开始时刻onsets和截止时刻offsets对应起来,采用一种更为准确的包络提取算法,并将信号包络进行onset/offset检测和匹配,通过时频标记得到目标语音片段和噪声干扰片段,将目标语音信号从混合信号中分离出来。实验结果表明:与谱减法和传统CASA算法相比较,该算法可提高目标语音的信噪比,降低算法的复杂度,改善算法的分离性能。更多还原

【Abstract】 Considering the shortcoming of instability and large calculation in segmentation of typical computational auditory scene analysis, a new speech separation algorithm based on onset and offset is proposed in this paper. The system matches onsets and offsets according to the same sources, adopts an accurate envelope extraction algorithm that can extract the signal onsets and offsets, tests and matches the selected onsets and offsets and obtain auditory segmentation, gets the target speech segment and noise fragment through the time-frequency marker, finally separates targeted speech signals from mixtures. The experimental results showed that compared with spectral subtraction and typical CASA algorithm, this improved algorithm can improve the segmental SNR of the target speech segmentation and improve the separation performance obviously.更多还原