Changelog [TODO] Train on VoxCeleb 2 dev set and test on VoxCeleb 1. Test attention component. Other pooling strategy. (Utterance-level Aggregation For Speaker Recognition In The Wild) Add multitask_v1 learning. [First version] Basic x-vector pipeline. Large margin softmax loss.