The goal of this paper is speaker recognition in broadcast news domain. The broadcast news audio is variable in length and accents, and also contains irrelevant signals, such as environment sounds. This paper proposes a two-path neural network to learn the embedding of speakers, one of which focuses on local (detail) features and the other focuses on global features. The features of the two networks will be merged, and new speaker embedding is proposed from these features. For the local network, this paper also aggregates its features by GhostVLAD algorithm. And GhostVLAD algorithm is formalized in this paper. The final result shows that Dual-vector has achieved remarkable results in the field of speaker recognition for broadcast news audio.