基于语音信号特性的沙漏网络下的人声与伴奏分离算法  

AN ALGORITHM FOR SEPARATING VOICE AND ACCOMPANIMENT IN HOURGLASS NETWORK BASED ON THE CHARACTERISTICS OF SPEECH SIGNAL

作　　者：孙超 余勤[1] 龚晓峰[1] 雒瑞森 Sun Chao;Yu Qin;Gong Xiaofeng;Luo Ruisen(College of Electrical Engineering,Sichuan University,Chengdu 610065,Sichuan,China)

机构地区：[1]四川大学电气工程学院,四川成都610065

出　　处：《计算机应用与软件》2023年第6期89-95,共7页Computer Applications and Software

基　　金：四川省科技厅项目(2020YFG0051);校企合作项目(17H1199,19H0355)。

摘　　要：沙漏神经网络最初是为了解决人体姿态估计而设计的,最近有的工作将其迁移到人声与伴奏分离的任务之中,然而这种网络结构较简单,并且分离性能较差,分离出的信号会产生伪影。为了进一步提高分离性能,针对语音信号的特性构造一种新的损失函数,可以使网络更好地学习和优化;在整个网络中加入批标准化和Leaky-ReLU激活函数,改进网络的反向梯度传播,稳定学习过程。在MIR-1K数据集上的实验结果表明,改进后的算法分离后的人声在源-失真比,源-算法引入伪像比上原始沙漏网络分别提高了0.18 dB和0.26 dB,分离后的伴奏在源-失真比、源-干扰比和源-算法引入伪像比上分别提高了0.23 dB和0.32 dB,明显超过了目前的人声伴奏的算法。The hourglass neural network was originally designed to solve the human body pose estimation.Recently,some work has migrated it to the task of separating the human voice from the accompaniment.However,this network has a simple structure and poor separation performance,and the separated signal will produce artifacts.In order to further improve the separation performance,a new loss function was constructed according to the characteristics of the voice signal,which could make the network learn and optimize better.Batch standardization and Leaky-ReLU activation function were added to the entire network to improve the network's inverse propagate and stabilize the learning process.The experimental results on the MIR-1K dataset show that compared with the original hourglass network,the improved network improve the separated voices source-to-distortion ratio and source-to-artifacts ratio by 0.18 dB and 0.26 dB,and the source-to-distortion ratio and the source-to-artifacts ratio of the separated accompaniment are improved by 0.23 dB and 0.32 dB.The separation performance of the algorithm exceeds the current vocal accompaniment algorithms.

关 键 词：沙漏神经网络 人声与伴奏分离 语音信号特性 反向传播 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]