收录:
摘要:
Aiming at the problem of insufficient feature learning of time-delay neural networks, which is widely used in the field of language identification, a new architecture called multi-scale and multi-dimensional convolution is proposed. The structure includes a global inter-frame correlation network, local and global multi-scale network, global channel correlation network, and multi-head attention statistics pooling layer. The global inter-frame correlation network models the global context at the initial frame layer to obtain the dependency characteristics of the global context, which makes up for the natural deficiency of time-delay neural network based on limited context; local and global multi-scale networks aggregate the information within and between layers to extract features on a finer and more complex scale; the global channel correlation network is explicitly modeled from the channel dimension to realize the adaptive correction of the channel dimension characteristics; The attention statistics pool layer is extended to multiple heads so that features can be distinguished from multiple aspects. Through the training of the AP17-OLR data set, it has been improved by 41% compared with the previous excellent model. © 2022 SPIE.
关键词:
通讯作者信息:
电子邮件地址:
来源 :
ISSN: 0277-786X
年份: 2022
卷: 12331
语种: 英文
归属院系: