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
温国伟 1刘金鹏 1方剑1
作者信息
- 1. 中国船舶集团有限公司第七二三研究所,江苏扬州 225101
- 折叠
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.
Abstract
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 sin-gle-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 memo-ry,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 mod-el.
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Key words
single channel/speech separation/deep learning/source-to-distortion ratio/source-to-ar-tifact ratio/source-to-interference ratio引用本文复制引用
出版年
2024
NETL
NSTL
万方数据