The xx205 System for the VoxCeleb Speaker Recognition Challenge 2020
This is an incremental improvement for speaker recognition systems, addressing performance in specific challenge tracks.
The paper tackled speaker recognition by developing a system for the VoxCeleb Speaker Recognition Challenge 2020, achieving second place with an equal error rate of 3.808% and minimum detection cost of 0.1958 in track 1, and 3.798% and 0.1942 in track 2.
This report describes the systems submitted to the first and second tracks of the VoxCeleb Speaker Recognition Challenge (VoxSRC) 2020, which ranked second in both tracks. Three key points of the system pipeline are explored: (1) investigating multiple CNN architectures including ResNet, Res2Net and dual path network (DPN) to extract the x-vectors, (2) using a composite angular margin softmax loss to train the speaker models, and (3) applying score normalization and system fusion to boost the performance. Measured on the VoxSRC-20 Eval set, the best submitted systems achieve an EER of $3.808\%$ and a MinDCF of $0.1958$ in the close-condition track 1, and an EER of $3.798\%$ and a MinDCF of $0.1942$ in the open-condition track 2, respectively.