摘要
在对音乐进行歌声与伴奏分离过程中,基于卷积编解码的分离模型虽然在一定程度上提升了分离效果,但网络存在丢失信息的问题。为了解决这个问题,提出一种有效的歌声与伴奏分离的模型。传统方法指的是非深度学习模型,本文在基于深度学习模型的基础上进行改进,通过在卷积编解码器的跳跃连接部分加入注意力机制,解决了网络丢失重要信息的问题,提高了分离性能。在开源数据集MUSDB18上进行验证,结果表明所提的模型分离效果良好。
In the process of separating singing and accompaniment of music, the separation model based on convolutional encoder-decoder improves the separation effect to a certain extent, but the network has the problem of losing information. In order to solve this problem,an effective model of separating singing from accompaniment is proposed. By adding attention mechanism to the skip connection part of convolutional encoder-decoder, the problem of losing important information in network is solved and the separation performance is improved.It is verified on the open source dataset MUSDB18, and the results show that the separation effect of the proposed model is good.
作者
王岚
WANG Lan(Communication University of China,Beijing 100024,China)
出处
《电声技术》
2022年第2期29-32,共4页
Audio Engineering
关键词
歌声与伴奏分离
卷积编解码器
注意力
separation of singing voice and accompaniment
convolutional encoder-decoder
attention