RES2NET SPEAKER IDENTIFICATION BASED ON CHANNEL GATING

Aimed at the problem that the current speaker recognition model has weak recognition ability to extract voiceprint features and cannot accurately identify the identity of the speaker, a speaker recognition algorithm based on channel gated Res2Net (CG-Res2Net) is proposed. A hierarchical residual connection was established in a residuals block by Res2Net to improve the voicing feature extraction capability of the system. The channel gating mechanism was used between the residual connection feature groups, and the important channels and relatively useless channels in the voiceprint feature were given higher and lower weights respectively. VoxCeleb1-test results show that the EER and minDCF of CG-Res2Net are better than those of Res2Net. Compared with ResNet network, EER and minDCF increases 38.05% and 17.95%, respectively, while compared with SE-Res2Net network, EER and minDCF increased 17.5% and 4.47%, respectively.