相邻位置编码Transformer网络:从卷积混合信号中估计源信号能量  

Neighboring Position Encoding Transformer Network:Estimating Energy of Source Signals From Convolutive Mixing Signal

作　　者：李腾辉 

机构地区：[1]广东工业大学自动化学院,广东广州510006

出　　处：《工业控制计算机》2022年第8期96-98,共3页Industrial Control Computer

基　　金：广东省重点领域研发计划(2019B010121001)。

摘　　要：提出一种基于Transformer结构的估计卷积混合信号中原始信号能量变化的算法。常用的Transformer结构使用sin和cos形式的位置编码,但是使用这种位置编码后其映射的特征不满足向量加法特性。在音频任务中不满足向量加法会对结果影响较大,为此,提出相邻位置编码,这种可训练的编码形式既能表示位置关系又满足向量加法特性。使用了这种编码方式的Transformer结构比原始结构能更好地估计原始音频信号的能量变化。This paper presents an Transformer based algorithm to estimate the energy of the source signal from the convolutive mixing signal.The commonly used Transformer takes the position encoding in the form of sin and cos.Nevertheless,this form of position encoding do not preserve the vector addition property of the embedded features.For this reason,this paper proposes neighboring position coding.This trainable encoding method can represent the position relationship and keep the vector addition property.In contrast with the position encoding with sin and cos,the neighboring coding method can have a better estimation of the energy change of the source audio.

关 键 词：卷积盲分离 TRANSFORMER 相邻位置编码 

分 类 号：TN911.7[电子电信—通信与信息系统]