一种基于自注意力机制的欠定盲信号提取算法

An Under-determined Blind Signal Extraction Algorithm based on Self-attention

自鸡尾酒会问题提出以来,欠定盲信号提取一直都是研究的热点和难点。应用较多的非负矩阵分解算法只能处理能以NMF建模的源信号。因此,利用深度网络替代NMF模型,将扩展多通道变分自动编码器与引入注意力的基于X-vector的目标信号判别模块结合,提出了一种欠定盲信号提取算法。使用VoxCeleb1数据集,通过RIR-Generate生成混合信号,利用BSS Eval工具对算法仿真结果进行分析评价。仿真结果表明,与基于MNMF的算法相比,提出的算法在欠定盲信号提取质量方面提升了2 dB,提取准确率提升了1%。

Since the problem of cocktail party was proposed,multi-channel underdetermined blind signal extraction has always been the hotspots and difficulties of research.While NMF methods work reasonably well for particular types of sound sources,one limitation is that they may fail to work for sources with spectrograms that do not comply with the NMF model.To address this limitation,deep neural networks are used to model the spectrograms of sources.This paper develops a sequential approach for blind source extraction by combining the generalized multi-channel variational autoencoder with the attention-based X-vector target signal discrimination module.It uses the Voxceleb1 dataset to generate a mixed signal test set through RIR-Generate,and the BSS Eval tool is used to evaluate the results.The simulation results indicate that compared with the algorithm based on MNMF,the algorithm proposed in this paper improves the quality of under-determined blind signal extraction by 2 dB,and the extraction accuracy is improved by 1%.