一种基于TCN的单通道语音分离算法  

A Single Channel Speech Separation Algorithm Based on TCN Network

作　　者：温国伟 刘金鹏 方剑 WEN Guowei;LIU Jinpeng;FANG Jian(The 723 Institute of CSSC,Yangzhou 225101,China)

机构地区：[1]中国船舶集团有限公司第七二三研究所,江苏扬州225101

出　　处：《舰船电子对抗》2024年第2期76-80,85,共6页Shipboard Electronic Countermeasure

摘　　要：针对端到端的语音分离中编码器模块仅仅使用一层的卷积神经网络,无法提取出语音信号更深层次的特征的问题,提出了基于时间卷积网络(TCN)的端到端的单通道语音分离,TCN有灵活的感受野,能够输入可变的序列,占用内存少,梯度稳定。利用中英文数据集对提出的基于TCN网络的端到端的单通道语音分离算法进行仿真实验,实验结果表明,与TasNet模型相比,所提算法的分离结果有所提升。The encoder module of the end-to-end speech separation only uses a layer of convolutional neural network,which cannot extract the deeper features of the speech signal.An end-to-end single-channel speech separation based on the temporal convolutional network(TCN)is proposed in this paper.The TCN has flexible receptive field,can input variable sequences,occupies less memory,and has stable gradient.The Chinese and English data sets are used to simulate the proposed end-to-end single-channel speech separation algorithm based on the TCN.The experimental results show that the proposed algorithm has improved separation results compared with the TasNet model.

关 键 词：单通道 语音分离 深度学习 信号失真比 信号伪影比 信号干扰比 

分 类 号：TN911[电子电信—通信与信息系统]